Energy Storage Management Using Artificial Intelligence to Maximize Polish Energy Market Profits

Abstract

Along with the growing renewable energy sources sector, energy storage will be necessary to stabilize the operation of weather-dependent sources and form the basis of a modern energy system. This article presents the possibilities of using energy storage in the energy market (day-ahead market and balancing market) in the current market conditions in Poland after reforming the balancing market in June 2024. The current state of the markets is characterized by high price volatility, which can ensure the high profitability of storage operations. However, very flexible and self-adaptive algorithms for charging and discharging are required, taking advantage of market price spreads. This study aimed to see if, through a solution based on ChatGPT 4o, energy storage operations can be planned by taking maximum advantage of the existing price spreads in the market. Previous analyses in this area have focused on complex models that predicted prices in the markets and planned the plant’s operation on this basis. In this case, the simple model used (charging and discharging based on historical prices) resulted in profits of EUR 90/MWh, while in the second case, when holidays, weather, and demand forecasts were taken into account, the profit was EUR 150–180/MWh, which exceeds the current Levelized Cost of Electricity of storage estimated at around EUR 100/MWh. These analyses indicated that modern genAI tools are appropriate for further study, especially as the technology dramatically increases its capabilities.

1. Introduction

As the rapid development of renewable energy sources (RESs) worldwide continues, energy storage facilities are beginning to play a crucial role in the new energy system. According to projections in the Renewables 2023 Analysis and Forecast to 2028 report, the global installed renewable capacity will reach 7300 GW by 2028 [1]. Energy storage’s main task is to stabilize power grids, which are increasingly dependent on variable sources such as solar and wind power. Energy storage facilities allow the storing of excess energy produced during periods of low demand so that it can be used when demand increases. This minimizes energy losses and increases the efficiency of the entire energy system. The introduction of these technologies is essential to achieving climate goals and ensuring the reliability of the energy supply in the future.

Poland is currently undergoing a rapid transition from coal-fired power generation to RESs. According to one of the scenarios developed in a report by the think tank Instrat, in 2040, the installed capacity of RESs in Poland will be 111 GW and energy storage of all kinds will be 27.3 GW. Currently, more than 7 GW of conditions for connection to the transmission grid have already been issued in Poland, according to data provided by the Polish Transmission Network Operator (TSO), and a significant number of such installations will be created at distribution networks. Comparing this with data for 2023, where the installed capacity of RESs was 28.6 GW with 1.8 GW of pumped-storage power plants [2], we can see what a rapid transformation awaits Poland and what challenges to ensure a stable system operation await us shortly. The current situation in Poland in the context of battery-based energy storage shows that this market de facto does not exist in Poland. According to a report by the Energy Regulatory Office, nine energy storage facilities with an installed capacity of 15 MW (data as of the end of 2023) are connected to Poland’s transmission and distribution network [3].

An example of a country/region that shows what role energy storage will play in the electricity system is California—the US state currently leading the storage market development—which is pushing hard to achieve a 90% reduction in CO₂ emissions by 2035. Currently, with a zero-carbon energy share of about 50% (30% from RESs, 10% from hydroelectric, and 10% from nuclear), it faces the challenge of creating an utterly zero-carbon energy system without developing nuclear power, which is not anticipated [4]. A significant component of this transition is the unprecedented development of solar power, combined with the development of energy storage, mainly in the form of battery storage.

California plans to achieve 100% zero-carbon by 2045, increasing the installed capacity from the current 35 GW to about 73 GW in 2030 and as much as 180 GW in 2045, mainly through the expansion of large-scale solar PV and energy storage facilities. These storage facilities are expected to play a crucial role in the new energy system, and California is a world leader in their use. Its energy storage capacity was 250 MW in 2019 and now exceeds 5000 MW, and is expected to reach nearly 20,000 MW by 2030, to reach 52,000 MW in 2045 [4].

Development Prospects of the Energy Storage Sector

The development of battery-based energy storage is inevitable. Currently, the largest storage facilities, such as the Vistra Energy Corporation (Sierra Drive Irving, Irving, TX, USA) in California (400 MW/1600 MWh) [5] and the Manatee Energy Storage Center (Parrish, FL, USA) (409 MW/900 MWh) [6], reach capacities comparable to medium-sized power units. It is worth noting that these storage facilities typically operate with discharge cycles of about 4 h. Changes in the market are occurring exponentially, with the available capacity and capability expected to double in the next five years. The global capacity planned in many countries’ systems is growing rapidly—although Europe seems to be behind on the impressive plans of California (as well as some Asian countries) for today—and so Germany will be around 35 GW in 2037 [7] and the UK 24 GW by 2030 [8]. It is worth remembering that Poland has already issued ca. 7 GW connection conditions for storage facilities. Storage facilities in Europe are also developing rapidly, although their capacity and storage capacities are somewhat smaller than those of US projects. Construction has begun on a massive battery energy storage facility (BESS) in Belgium. The installation received a construction permit a year ago and won a capacity auction in the fall of 2023. The planned power and capacity of the BESS will be 200 MW/800 MWh [9]. In Poland, preparations are underway to finalize a tender for Europe’s largest energy storage facility located in Żarnowiec—with a capacity of up to 263 MW and a minimum capacity of 900 MWh [10]. The final cost-effectiveness of the construction and operation of these storage facilities is widely debated. They usually try to combine different revenue models—participation in the capacity market (here, it is worth noting the large share of storage facilities in the Polish market), direct operation using the daily price spread (and co-daily discharge cycles), and also assistance in direct frequency regulation for transmission systems (here battery storage facilities with a response time of about 0.02 s are unbeatable). Pricing data at present are wildly divergent, but it can be expected that in the near-term the cost analogous to the Levelized Cost of Electricity (LCOE) will be (optimistically) at the level of USD 100/MWh (although some reports today show a figure twice as high [11]) or even additionally show a substantial decline by 2035 (which shows on what spreads storage can operate), and the available reports from California already show cost-effective operation at USD 70–100 per year per kW of installed capacity [12]. The technology is not a niche and is not on heavily subsidized terms. It is realistically entering the stage of commercial development. There is a rapid increase in available capacity, the capacity of storage factories is growing tremendously, the price is falling, and the construction and installation itself is already a standard process with an extremely short (for an energy system) timescale. The storage business is also correlated with electromobility (analogous battery production technologies), which will substantially impact cost reduction. It is already a foregone conclusion that this will be one of the growing markets [13].

The capital and operating cost of energy storage depends on many factors, including storage technology, installation scale, and location. The average investment cost for battery storage, such as lithium-ion, currently ranges from about USD 300 to USD 500/kWh of installed capacity. Operating costs primarily include maintenance, battery replacement, and energy consumption during charging and discharging. The total operating cost can range from USD 10 to USD 20/MWh.

For an energy storage facility to be profitable, its revenues must exceed the sum of the investment and operating costs. Current analyses indicate that an energy storage facility must generate profits of about USD 100–150/MWh to cover both its investment and operating costs and provide an acceptable return on investment. For example, a study published in Batteries (2023, vol. 9, no. 3) indicated that with the proper charging and discharging strategy, it is possible to achieve profitability even at lower costs, mainly when the storage is used within the day-ahead market, where energy prices can be more favorable [14].

Regardless of the plans and subsequent stages of the energy transition, the technology that develops most rapidly and cost-effectively will be decisive in Poland. Energy storage facilities can become a quick alternative to changes in the energy system, emerging naturally from current capacity market contracts. If even half of the planned 7 GW of storage facilities can be realized by 2030, their impact on the power system will be significant, and their operation will gradually begin to displace other generation units, such as coal-fired power plants. They could also pose a serious challenge to future gas projects, which are already under pressure from the European Union’s ambitious climate targets and increasingly stringent emissions reduction regulations.

However, it is important to remember that, unlike in California, battery energy storage will not be a universal solution in Poland. Our specific weather conditions, especially in winter, cause difficulties that require a more comprehensive approach. While increasing the number of wind farms, including offshore projects, is a step in the right direction, there is still a risk of prolonged periods without sufficient production from RESs, which can last for weeks, especially in cold weather. For this reason, the Polish energy system will have to rely on more long-term storage, which is currently yet to be commercially fully developed, or reserve units, where gas will still play a key role.

The need to ensure the supply of system heat for industry may encourage the greater use of gas systems, such as steam–gas power plants. In this way, Poland can return to the concept of transformation based on the RES-storage–gas model, where gas, over time, will be able to evolve into a hydrogen economy. However, this technology’s development is lagging behind earlier projections.

2. Literature Overview and Market Operation

Energy storage facilities have operated in electric power systems worldwide for many years. The authors in Energy Strategy Reviews 54 propose the most straightforward division, distinguishing thermal, electrical, electrochemical, chemical, and mechanical storage. Within this article, the authors also raise an important issue related to the number of related publications (more than 300 articles and more than 50,000 citations in 2021), which indicates a significant increase since 2006 in all categories, with the most analyzed area being battery energy storage [15]. The type of storage is directly related to how it is used. The authors in Electric Power Systems Research 236 (2024) pointed out that technologies such as pumped-storage (PHS) or power to H₂ power plants perform better at discharging over a longer horizon (hours, days, and weeks), while battery energy storage works well for times ranging from seconds to days. This allows them to be used for frequency control, load management, and energy market operations [16]. The authors of Renewable and Sustainable Energy Reviews 148 (2021) point out that 91.9% of installed capacity is PHS, while electrochemical storage is only 5.9% [17]. However, juxtaposing this with information on planned projects and current research, it should be pointed out that shortly, battery energy storage will be the central area of development of energy storage technology. Thanks to their flexibility, electricity storage facilities are distinguished by a wide range of possible uses like price arbitrage, capacity market participation, and ancillary services. The storage facility can also stabilize the distribution grid and cooperate with a renewable energy facility to become a comprehensive energy generator or a source of energy for industrial facilities.

In the current phase of market development, storage viability and investment decision making are based on capacity market auctions. These were introduced in Poland in connection with the need to ensure the stable operation of the National Power System by ensuring the continuity of electricity supply to end consumers. A distinction can be made between generating units, which undertake to supply power, and consumption units, which undertake to reduce consumption. These activities take place when the Transmission System Operator announces a recall period. Units participating in the capacity market receive standby remuneration. The amount of remuneration is determined in annual power auctions, which apply to a delivery period starting five years after the auction and lasting up to 17 years [18]. From the perspective of an energy storage facility, the capacity market is a stable revenue base, which, however, does not fully allow it to ensure a return on investment.

With the radical decline in storage investment costs and the increasing daily spread in wholesale energy market prices or new opportunities in the balancing market, at a rapid pace, storage facilities may become profitable only by operating in the form of price arbitrage (charge–discharge) in daily cycles or by providing balancing services for energy and balancing capacity.

In June 2024, there was a new revenue opportunity for energy storage facilities in Poland, and that is participation in the provision of balancing services to the Transmission System Operator [19]. Balancing services consist of balancing energy and balancing capacity. In the case of balancing energy, these are activated when energy production and consumption do not equal each other. Balancing capacities are capacity reserves that are available in the power system to respond quickly to deviations from balance. Balancing capacities fall into three main categories:

The FCR (Frequency Containment Reserve) is a reserve that is automatically activated in response to frequency fluctuations in the network, so as to keep it within an acceptable range. The FCR is crucial in the first seconds after a disturbance occurs, as shown in

Table 1. Execution time required by the TSO.

Service	Execution Time
RR	≤30 min
mFRR	≤12.5 min
aFRR	≤5 min
FCR	≤30 s

; for this reason, it is essential to have the right IT systems to automatically activate the reserve [20].

The FRR (Frequency Restoration Reserve) is a reserve activated after the FCR to return the system frequency to its nominal value. It operates on a minute scale, and we can distinguish between an automatic version (aFRR) activated by control systems in response to transmitted signals from the TSO, and a manual one (mFRR) activated manually by the installation operator.

The RR (Replacement Reserve) is activated to replace previously activated FRRs and ensure long-term system stability. The RR is activated on a scale of minutes to hours and is used to restore the power reserves used by FRRs, so that they can be available again for quick use when needed [21].

A catalog of the above services has emerged with balancing market reform. Due to the limited number of entities, the services currently provided are expensive. However, the balancing market is a technical market where the Operator manages energy supply and demand to maintain adequate frequency parameters in the grid. The characteristics of the market, including the method of price forecasting, can take time due to the multitude of factors influencing the final price in the market. This is related to, among other things, the currently available generation units, their type, the share of RESs, and the current demand in the grid. Poland is also interconnected with its neighbors, so the current energy situation in these countries may affect the final price.

Two factors are critical. The rapid decline in investment costs—according to IRENA projections, some electricity storage technologies will be 66% cheaper in 2030 compared to 2016 [22], which is also related to the rapid development of the electromobility market, and systemic changes in modern energy markets and energy systems introducing RES capacity exponentially. Due to the problem of weather dependence on wind, as well as solar sources in many energy systems, we have to deal with large daily fluctuations in wholesale prices in spot markets—from even the “classic” phenomenon of “duck neck”—lowering prices in the middle of the day due to solar generation, a significant increase in energy prices during the peak, to price fluctuations between individual days with good and bad wind conditions (and associated generation from wind farms). Storage facilities operating on price arbitrage thus have an excellent field for realizing profitability and, at the same time, have a positive effect on the energy system—which encourages individual countries to subsidize their development through capacity markets. Price volatility in the wholesale and balancing markets is very high.

Prices on the wholesale market are characterized by high volatility (especially in the summer months and from June 2024—the change in the balancing market (BM) price determination system in Poland). The market changes (pricing rules, modification of billing from 1 h to 15 min, a rapid shift in generation structure) cause difficulties for typical modeling and price forecasting (historical data is mainly inadequate for the current state of the market). Therefore, it seems adequate to use the proposed ChatGPT modeling. A paper presented the results of the possibility of obtaining revenue from storage operation in the winter months (February 2024)—where the potential income from storage operation was at the level of about 50 EUR/MWh (more than two times less than currently in the summer season) [23]. However, the price level in the future is unpredictable (it is worth referring to the enormous price increases in the markets after the start of the war in Ukraine).

Figure 1 and Figure 2 show the price volatility on the balancing and exchange markets (the Polish Power Exchange’s day-ahead market (POLPX DAM) segments) in June 2024. This month, the rules of the BM were changed—15 min intervals for determining prices were introduced, and the way of determining prices was modified, which was supposed to allow the market to operate better. All this happens in the summer months when there is a lot of RES generation (photovoltaics), which causes additional price fluctuations (among other things, the RB is supposed to counteract the mandatory switching off of RES generation when there is an excess of generation, which is becoming a common occurrence in Poland).

Strong price fluctuations can be seen (1 EUR = 4.3 PLN), reaching maximums above 1500 PLN/MWh. Negative prices have also commonly appeared. All this leads to favorable economic conditions for the operation of energy storage facilities.

Daily price volatility results in a price spread (min–max price difference during the day shown in Figure 3) of up to 1400 EUR/MWh, which can be used in storage operations. The maximum potential price spread is about 300 EUR/MWh, and the minimum is 70 EUR/MWh (POLPX DAM) and about 0 EUR/MWh (BM). Potential profits can be huge.

With the simulated storage operation in a daily cycle (1 h charging, 4 h discharging), the optimal choice of markets, and without considering the efficiency of charging/discharging, it is possible to achieve high profits at levels (monthly averages) of POLPX DAM—160 EUR/MWh, BM—155 EUR/MWh, optimal strategy with market selection—even 180 EUR/MWh, and this is summarized in the Figure 4. These values could even be increased if the storage operated in double daily cycles—in such a simulated case (operation on the BM), average profits per day could reach up to 290 EUR/MWh (Figure 5). At this moment in the simulation, the BESS efficiency was neglected.

The only problem, then, becomes the choice of charging and discharging hours for storage and the choice of the most profitable market for the storage to conduct price arbitrage (use the spread). The forecasting of generation from a wind or solar farm itself is a task that is quite well mastered scientifically and commercially through the use of physical models of the installations, statistical modeling, and forecasting using advanced machine learning techniques. Examples of error values for forecasts in recent years are that between 2005 and 2019, the Weighted Mean Absolute Percentage Error (WMAPE) for wind farms in Germany changed from 37.2% to 11.9% [24], and for offshore wind farms, the Mean Absolute Percentage Error (MAPE) in the 2024 models was already 2.5% [25]. WMAPE values of 4–5% have been emerging for photovoltaic farms in recent years [26]. Thus, it is possible to predict the operating profile of a wind or photovoltaic farm (which positively solves the problem of integrating storage with RES production in farms). Forecasting prices on the DAM or BM itself is much more difficult. There are sophisticated price-forecasting methods that sometimes show promising results, but building such models into systems for optimizing energy storage operations is a rather complicated and often expensive assignment. There are researches like the one in the Ain Shams Engineering Journal 15 (2024), which explores the application of machine learning and deep learning algorithms, specifically artificial neural networks, adaptive neuro-fuzzy inference systems, long short-term memory, and gated recurrent units, for accurately forecasting short-term electricity prices. It also shows that incorporating external factors, like temperature and calendar data, significantly enhances prediction accuracy [27]. The article in Utilities Policy 70 presents an ensemble learning approach for very short-term electricity price forecasting in markets with renewable energy sources, combining techniques like extreme gradient boosting and random forest with Bayesian linear regression as a meta-regressor [28]. In Germany, a study included an ensemble approach to enhance day-ahead electricity price forecasting by aggregating multiple machine learning models, improving the accuracy in capturing the dynamic conditions of the energy market [29]. There was also a study that focused on prices in the balancing market. During their research, they checked a variety of statistical, machine learning, and deep learning models, and the conclusion was that simpler statistical models outperformed complex ones [30]. As you can see, there are sophisticated price-forecasting methods that sometimes show promising results, but building such models into systems for optimizing energy storage operations is a rather complicated and often expensive assignment and usually well researched.

The storage could operate effectively and economically on relatively simple operating models (selection of charging and discharging hours). However, it should be quickly adapted to new conditions (grid problems, cross-border flows in the market, sudden weather events, failures, etc.). The proposed solution is to use Gen AI (Chat GPT model 4o) to support the working strategy of the storage operating in price arbitrage in the wholesale and balancing market.

Analyzing the data from the new balancing market in Poland, one can see that it is becoming challenging to forecast prices accurately. Analogously, in the summer months, this happens in the SPOT market. This is all due to the current generation structure, where, when solar conditions are good, PV sources push other power plants out of the market (and reduce the price quite significantly). However, simultaneously, with the end of their operation in the daily cycle, periods of high prices appear (coal and gas sources). On top of all this, there are also weather conditions, such as wind and variability in wind farm operation. However, it should be noted that such a high price volatility is a feature of the current summer months and is influenced primarily by weather conditions and new market regulations (BM). Indeed, daily price spreads on the Polish market in the winter months (e.g., February 2024) were much smaller and oscillated around 50 EUR/MWh maximum.

Given the many possible uses of energy storage and the significant price dynamics in energy markets, it is necessary to provide tools that can autonomously plan the operation of plants. So far, the power industry has been based on large power units, where the changeover occurred at times longer than in energy storage. Current trends indicate the progressive development of distributed energy and the emergence of many installations with smaller installed capacity, which translates into limited revenues. Owners of such installations need a solution that is simple and inexpensive to implement instead of using complex computational models.

Use of Artificial Intelligence in Energy Storage Operation

The emergence of profit opportunities in storage operations needs to translate into the choice of a single strategy for storage operations. One thing to analyze post factum is historical prices, and another is to optimize storage operation realistically (which would require forecasts). Of course, one possibility is to select (by statistical analysis, also with the help of AI) typical hours with low prices (charging) and peak prices (discharging). Then, it is possible to create models for weekdays and holidays, finally supplemented with weather data (wind, sun, and temperature forecasts) and using the system demand forecast (available on the Transmission System Operator’s website). In the final solution, the model will be supplemented with data on the availability of generation sources (REMIT). It is also possible to use any data—especially local data—which will undoubtedly contribute to a better accuracy in strategy selection.

ChatGPT appears in several studies, but this is only in the preliminary stage; the studies refer to possibilities, and most focus on already known AI solutions. In the Journal of Advances in Artificial Intelligence, several applications for Large Language Models (LLMs) are mentioned, including support at the stage of sizing storage for RES installations, analysis of the market situation, price estimation, or analysis of large data sets for potential optimization [31]. The use of other artificial intelligence solutions is also indicated in the Energy and AI 14 (2023) article, which presents a novel deep learning with reinforcement (DRL) approach for optimizing the coordination between wind energy and battery energy storage systems (BESSs) in wholesale energy and ancillary services markets. The proposed strategy increases financial returns by reducing wind constraints and maximizing revenues through dynamic bidding in the common market, significantly outperforming traditional optimization-based methods [32]. Smart strategies for BESSs operating in microgrids are also analyzed. The authors compare heuristic methods with advanced artificial intelligence (AI) methods such as fuzzy logic and artificial neural networks [33].

As the analysis indicates, energy storage facilities are necessary to stabilize the power grid in the future. The plethora of uses drives storage operators toward artificial intelligence-based IT solutions. A novelty is the use of ChatGPT proposed in this research, which has not been used in energy storage scheduling so far.

3. Materials and Methods

The analyses below use simple AI models using Chat GPT 4o, considering only data from the learning set (June 2024) tested on the validation set (BM and POLPX DAM prices July 2024). Using generative artificial intelligence, one can rely on one’s models or use available data analysis models from Data Analytics Wolfram, among others. Both of these approaches were used in this case. The use of GenAI means that in the analyzed case, the model is not in explicit form (nor practically in mathematical form); only the answer is given. A significant advantage of such a solution is the ease of data input—the model can interpret the data provided and take them into account later in the results presented. GenAI optimally prepares the storage schedule. Under current market conditions, if the storage facility is to operate using the price spread (difference) on the DAM markets of the power exchange or the balancing market, the model is not required to operate online (the market structure needs a “day-ahead” bid). Thus, the modeling itself and the result are different from the use of AI models in the past for the ongoing optimization of equipment operation—as they were used by the author in many works—including an immune-inspired optimizer of combustion processes in a power boiler; modeling solid oxide fuel cell behaviors by an artificial neural network; and a eat demand forecasting algorithm for a Warsaw district heating network [34,35,36]. At that time, the models (usually a combination of NARMAX models—linear with the addition of a multilayer perceptron neural network or various types of neural network structures, also in combination with fuzzy logic systems) generated current-operating points (or directly predicted demand levels or other technical data, which were then the basis for predictive control). They were created based on quite a labor-intensive and sophisticated data analysis. In the operation of the storage facility, the optimal solution can be boiled down to the selection of charging and discharging hours (precise forecasting of market prices is not required, only the scheduling of the storage facility operation, which must work in daily cycles).

For this reason, the operation of the ChatGPT (model) is focused on the selection of optimal charging (lowest price) as well as discharging (highest) hours, and the solution to the price modeling problem is reduced to the determination of the most probable hours with minimum and maximum price levels. In the case of highly volatile markets (and this is what we are dealing with) and the variability in prices due to day types, weather conditions, and also seasonality (summer–winter and the dominant influence of PV generation on the price profile), it is most desirable to create a tool that could give the optimal hours with reasonably good accuracy (but not necessarily online), one that allowed the rapid adaptation of the model in the event of the emergence of new operating conditions. The ChatGPT model is optimal for this purpose. We are easily able to feed the model with data (any file format or GPT API) and then obtain the results (optimal working hours of storage), and then control and modify the model (or even update it every day). It is worth realizing the labor-intensive nature of the typical solution so far (price forecasting) in the case of a modern balancing market (15 min prices, high volatility), when the only thing sought in practice is the selection of the proper operating hours.

GenAI systems like ChatGPT are precisely the perfect solution for getting a quick solution to highly variable external factors. At the same time, the current level of model availability (ChatGPT) is not a flawless solution. AI systems are focused on obtaining and giving a result under all circumstances, without careful control of the result obtained. Thus, we may receive a suboptimal and even far from the optimal solution. However, in the case of storage scheduling, the results given by the AI can be compared to the performance of an “expert trader” who uses his market experience to approximate market forecasts and historical price correlations with external conditions.

The modeling used training data from June 2024 for the balancing market (Figure 6) and the power exchange (DAM segment), along with additional data (demand forecast, RESs) available on the Polish System Operator’s website. They were tested on a test set—data from July 2024. The results were presented for 1 MW of storage capacity without considering storage efficiency.

4. Results

The purpose of this study is to preliminarily determine the applicability of GenAI models in scheduling production from battery energy storage under current market conditions. In particular, determining the potential profits (EUR/MWh) for the storage operation in daily cycles can be a basic investment guideline for storage construction. It is possible to extend the models extensively, taking into account revenues, e.g., from capacity markets (in which case part of the storage capacity is reserved (but yielding guaranteed revenues)) or more advanced storage operation strategies (such as multiple charging and discharging per day—Figure 7). This article assumes 100% storage capacity utilization under single-cycle storage conditions (Figure 7 and Figure 8). The advantage of ChatGPT scheduling is that results are quickly obtained when different strategies are adopted.

Determining the hours with minimum and maximum price levels based solely on historical data is subject to a large degree of uncertainty (as can be seen in the results of Figure 8—the possibility of using the maximum spread). This is, of course, due to the volatility of prices (especially the balancing market), where prices are derived not only from typical daily demand profiles (which, in turn, are variable seasonally and also on different types of day) but also from the current situation in the energy market—among other things, the demand–generation relationship (which is derived, among other things, from weather conditions and the operation of RES sources—the production of PV or wind sources). In the simplest models, it is possible to determine, for example, typical hours of storage discharge (the highest price levels per day)—shown in Figure 8 (here as averaged hours of the highest prices over the entire modeling period without the influence of additional external factors). Much better results can be obtained if more accurate models are prepared (separately for holidays and weekdays—which amounts to modeling different demand profiles), as well as the use of other external data. (Here, demand forecasts from the Transmission System Operator were used, as well as weather and RES generation forecasts. In such cases, the ChatGPT model modified the forecast hours (on each day).)

In the modeling, the energy and balancing markets were analyzed (in the current study as hourly averages, with an optimization of the work for the 15 min market in preparation). In the most straightforward solution—only price analysis (post factum) and the application of scheduling proposed by the ChatGPT model—the result was (Figure 9) that the storage facility can obtain revenues at 40% of the spread.

A simple ChatGPT model (selection of charging and discharging hours) allows for a hypothetical revenue of about 90 EUR/MWh with a maximum potential of 170–190 EUR/MWh (one-time charging/discharging cycle per day) (Figure 9). In further work, the model was adjusted for weekdays and holidays and supplemented with demand and weather forecasts (available on the pages of the system operator). Adjustments were made to the “day ahead” schedule—generating revised operating schedules (the modification of predicted optimal charging and discharging hours), which allowed an increase in the projected revenue to a level of about 60%; the schedule is shown in Figure 7 and the achieved results in Figure 10.

IT Systems Needed to Operate in the Market—Real Case Scenario

The model alone will not be able to function in the energy market. Specific IT solutions are needed, some of which are clearly defined by the Transmission Network Operator and some of which are used independently of the markets.

Energy storage consists of more than just a set of batteries and a primary control. In the context of operating in many markets, it is essential to have a system that allows control signals to be managed by prioritizing them appropriately and passing them on to the Battery Management System (BMS). This solution is called an Energy Management System (EMS), and it is used in many places. However, it should be noted that this is not only an IT solution but also a real-time controller that can meet the Operator’s requirements in terms of response times for the services provided, which, as shown in the Table 1, can be as low as less than 30 s. It is also worth noting that the Transmission System Operator, in its published standards, indicates that the entire communication loop between the Operator and the energy storage can be a maximum of 5 s.

Considering the regulations for the operation of the balancing market, the Transmission Operator has identified three IT systems necessary for communication between the TSO and the ESS.

From the perspective of the technical operation of energy storage, the system that does not directly affect the control of the ESS is WIRE—the Energy Market Information Exchange System. It is responsible for exchanging commercial and technical information about the energy market in Poland. Information on concluded energy sales contracts is sent through this system, and in the case of the balancing market, users submit bids for energy and balancing capacity through this system. The document that can affect the technical operation of the storage facility is the work program, which is a document in which the Balancing Services Provider (BSP), the party that is responsible for the storage facility to the TSO, transmits the energy storage facility’s work plan in 15 min resolution.

The System for Operational Cooperation with the Power Plant is also responsible for two critical tasks when communicating with the TSO. Through SOWE, the availability of a given unit is reconciled, and the ESS Operator must report all planned and unplanned outages, failures, and maintenance. With the information thus provided, the Operator can plan the National Power Grid operation. In the other direction, the Operator sends the Current Daily Coordination Plan through the SOWE. It forms the basis that the ESS’s EMS system must consider when sending signals to the various ESS components. The SOWE provides a backup communication channel if the system also operates an LFC System. The LFC or Load Frequency Control is the system responsible for the control of units by the TSO and consists of two modules. The first is responsible for transmitting signals by the Operator regarding the plant’s current operation, including those related to activating individual services for balancing capacity, such as the FCR, aFRR, mFRR, or RR. The second module is the System for Monitoring Operating Parameters (SMPP), which is responsible for sending to the TSO data on the current operation of the installation, such as net and gross power, information on activated services, or, in the case of renewable energy sources, estimates, i.e., the maximum power that the installation could generate under current weather conditions. Based on the data transmitted by the SMPP, the Operator has control over whether the installation is operating according to plan and can verify whether the response to individual signals for services is based on models prepared by the TSO. Since the LFC requires close cooperation with the unit, it is necessary for it to communicate it with the EMS system and automation systems located at the facility. The TSO in its standards indicates that if the unit has an LFC system, any TSO signals must automatically and immediately be executed with feedback.

The system affecting all components of the energy storage will be an optimizer. This term is a tool whose task will be to plan the optimal operation of the storage unit in several markets based on the implemented algorithms. However, it should be remembered that, in many cases, the TSO makes the final decision on the unit’s operation plan. In this publication, we introduce the possibility of using ChatGPT to replace existing IT solutions. Fewer errors characterize complex models. However, their cost is also significant, which, in the case of small storage installations, can cause problems. The whole set of products which are essential is presented in Figure 11.

As can be seen, for an energy storage facility to function correctly in the market, a broad set of necessary information systems is needed for communication with the TSO or energy trading. The optimizer under analysis in this article is attached to the system of necessary systems by integrating them. It uses various external data related to historical prices or forecasts. Work on the target industrial solution is underway so that the system will be ready when the ESS joins the balancing market (currently waiting for the first such project in Poland).

5. Discussion

Energy markets are entering periods of high price volatility. Recent summer months show a spread (both in the balancing market and the exchange DAM) of 300 EUR/MWh (monthly averaged daily values—Figure 3) and maximum values that can reach up to 500 EUR/MWh with the emergence of negative prices.

Even in the case of simple storage utilization strategies (single daily charging and discharging cycles), there is a high potential for revenue gains using the scheduling of operations with GenAI at a level of up to 60% of the daily spread, i.e., about 150–180 EUR/MWh, which, taking into account the efficiency of storage facilities, gives final values exceeding current cost levels (LCOE)—which in most studies are presented at the level of about 100 EUR/MWh. Over the next period (1–2 years), one can expect a significant decrease in the cost of BESSs (projected even to the level of 50 EUR/MWh), which will lead to the guaranteed profitability of storage operations even with more minor differences in wholesale prices (as was the case in the winter months in Poland).

Interestingly, the obtained simulations of profits from a storage operation in a double daily cycle were lower than the optimal solution for a single daily cycle (a revenue simulation result of only 73% of the result (profit) for the model modified for a single cycle was obtained)—the schedule is shown in Figure 12. However, it is not easy to conclude that using single daily cycles of storage operation is more profitable—this, of course, depends on the nature of the price distribution. The models can then be used extensively in the variant analysis of successive strategies, taking into account the possibility of single or double daily operation cycles, variant lengths of discharge time, variant uses of the balancing market or power exchange (or both markets simultaneously), and the possibility of the operation of the storage facility with a partial power reservation (capacity market revenues). The optimal strategy of storage facility operation can also change in successive months (autumn or winter season). The key is easily using ChatGPT models and quickly analyzing variant strategies.

This publication shows that relying on statistical indications of optimal charging and discharging points may only sometimes result in the correct outcome. If energy prices vary in a disorderly way and at times other than those statistically predicted, such a simple model may have problems estimating this correctly. This analysis also showed that a profitable model for a storage facility could be based on publicly available data provided by the Transmission Provider. Here, too, there is room for a potential expansion of the study to include more complex forecasting models than those from which data are made available by the Operator based on classic AI methods. By combining forecasting through advanced algorithms with an analysis of these data by new language models, we will obtain significantly better results. As this is the initial analysis stage, the model does not consider many technical aspects of storage operation, including plant efficiency or more complex market conditions related to the Operator’s control of the plant at times other than the plant owner’s plans.

6. Conclusions

Battery storage is the future of energy systems. It should not be discussed; it should only be rapidly invested in and, therefore, change how power systems operate. The example of Poland—a country with a historically huge share of coal in the generation of electricity (a few years ago still 75%—today falling to around 50%, and eliminated from the energy system around 2035)—shows this clearly, with a rapid investment in RESs (even up to 50% in 2030) and tremendous investments in energy storage (a tender for the largest storage in Europe, 7 GW of permits for the construction of new storage already).

The operation of storage facilities will become more and more profitable—in Poland, this has been shown in recent months (summer months). It is becoming possible to achieve income from storage operations at the level of more than 100 EUR/MWh—which, with the current cost levels and looking at the falling prices of storage facilities, ensures the profitability of the investment [12,23].

Of course, the targeted deployment of large storage capacity will automatically reduce the price divergence (which is right for the system, less so for storage investors). However, at the same time, investment prices will fall, resulting in favorable conditions for deploying storage capacity.

Here, we are remodeling our power industry towards a new energy system—like California, where demand will be covered a 100% by renewable sources working with energy storage. In Poland, the transition process will undoubtedly be more difficult partly due to the weather problem, and the system will, for a long time, use reserve sources working on fossil fuels (gas, leftover coal). However, it seems that the direction of the transition is a foregone conclusion.

One should expect even more volatility in the markets and the widespread appearance of negative prices, but at the same time, significant price fluctuations—even in 15 min intervals—make it even more challenging to plan storage operations optimally [19]. Specifically, negative prices and momentary volatility can only be considered in AI models. The demand for fast and straightforward AI models will increase with the number of distributed RES installations.

Artificial intelligence systems (such as ChatGPT) in the years to come, with the increase in their computational and inference capabilities, and most importantly through their ease of deployment configuration, could be a very interesting solution for selecting the optimal energy storage operation strategy [31,33]. This publication was the first to use ChatGPT to carry out simple strategies for energy storage operation, which may change the current approach of using complex solutions in favor of simpler systems; so, it is recommended to continue the analysis in the area of the application of ChatGPT working with energy storage. Further research will be conducted to improve the quality and certainty of the results obtained by using more data and better-quality predictions. In parallel, the model development area should be studied to develop resilience to unexpected energy market behavior.