Using Artificial Intelligence to Predict Power Demand in Small Power Grids—Problem Analysis as a Method to Limit Carbon Dioxide Emissions

Abstract

The article discusses the application of advanced data mining methods applicable to electricity consumption within a local power system in Poland. This analysis involves power demand. It is aimed at predicting daily demand variations. In such a case, system demand is characterized by high variability over a short period of time, e.g., 24 h. This constitutes a significant issue within a small power grid. It entails effective load programming on a given day and time. Therefore, the authors of the paper suggested employing artificial intelligence to forecast industrial power grid load for successive time intervals of the operation process. Such a solution applied within a power system enables appropriate start-up/shut-down planning, as well as generator operation at a specific capacity in power plants. It thus allows continuous power system (on-line) load demand balancing. Predicting power system load also involves determining moments, e.g., of power plant start-up, transition times to maximum or minimum output, or also the shut-down of such a process. This means ongoing and continuous (automatic) impact on electricity distribution. It significantly reduces carbon dioxide atmospheric emissions and allows zero-emission, e.g., wind, hydro, geothermal, or solar plants to meet current power needs. The issue associated with operating small ‘island’ power systems is a dynamic and rapid change in power demand. This is related to the area-based—‘island’—use’ of available power sources that can only be operated within a specific area. A very important problem occurring within these structurally small grids is the continuous forecasting of load changes and real-time response to power demand (i.e., balancing power demand through in-house or available power sources).

1. Introduction

The power grid of a specific country is a combination of conduits, overhead or underground electric tractions, or other functionally related equipment that converts electricity, e.g., transformers and switching stations, etc. [1,2,3]. All these components or devices are electrically coupled, which means they form an always-oriented connection network with specific, assigned relations, e.g., power plant—high-voltage line—transformer station—power consumer [4,5,6,7]. National power systems are intended to transmit (power lines, PL), convert (transformer stations, TS), and distribute (switching stations, SS) electricity generated in emission-generating and zero-emission power plants (Figure 1) over a specific area. Electricity is sent to all users, i.e., consumers, e.g., residents, industrial plants, buildings, and state critical infrastructure facilities (SCIF), etc. [8,9,10,11,12]. Electricity generated in power plants is used to power all electrical and electronic consumers. It is essential for their operation and implementation of all programmed functionalities. The above components and devices comprise a specific power system employed within a given country and classified as SCIF [13,14,15,16,17,18,19]. A power system includes specific elements (such as PL, TS, or DS) that function as a whole. Those are interrelated elements. In other words, specific functional relations, e.g., TS—SS, PL—TS, etc., and feedback can be allotted within a power system. A power system (PS) is always internally and externally coordinated when interconnecting with a different national transmission system. A PS always forms a specific technical structure of connections, as well as reliability. It is a set of components and devices used online to transmit and distribute electricity. To improve electricity transmission reliability, a PS always includes an additional set of elements ensuring the so-called redundancy. Redundancy can be hot (operating at a given moment) or cold (planned for start-up due to a failure) [4,20,21,22,23,24,25,26].

The following power plant types can be distinguished within a national power system—renewable and non-renewable; employing conversion to thermal energy within their process; as shown in Figure 1. Climate-related threats that are a current global challenge (ozone hole, climate change, etc.) drive the development of power plants classified as zero-emission, e.g., on- and off-shore wind farms, hydro, geothermal, and solar power plants—Figure 1. Power plants fired by non-renewable sources include two types of power facilities that utilize different materials to generate required heat and energy, namely, most usually coal and fissile fuel—Figure 1. A noteworthy parameter that characterizes all of the aforementioned power plants operating within a PS is the so-called inertia time (defined in electronic engineering as a time constant τ) of the response to power demand at a given moment [19,27,28,29,30,31,32,33,34]. It is a particularly important technical parameter that is employed to balance the power demand for a given PS, and at a specific moment in time, e.g., start-up, synchronization, connecting a different generator to the grid, and also reaching specific available capacity within the system required to meet the entire demand, etc.

Figure 2 is a general illustration of a PS, including interlinks and relations between ST, SR, and LE. Due to the fact that they are being used to convert electricity, the O₁, O₂, O₃, …, O_n consumers are characterized by different load parameters, including current consumption at time t. O consumers are located at geographically different sites within a given PS—Figure 2. There are also usually time-varying and variable (pulsed or constant) power demands of such equipment.

Electricity generated in a power plant by a high-voltage TS (A) is supplied to an overhead power grid. Via high- (C), medium- (D), or low-voltage (D) PL, it is then fed via a TS (B) to electricity consumers that usually exhibit a varying rated capacity: very high MW (10⁶ W), medium kW (10³ W), as well as low (W). Start-up (incorporation) characteristics, including specific consumers’ power demand, differ, as shown in Figure 2 [24,27,35,36]. They vary from pulsed current waveforms at an initial start-up time O₁ to steady. Most often, other consumers have set power demand waveforms. Only pulsed changes, variable over demand time, most often appear at consumer start-up time O₂, O₃, …, O_n [11,26,37,38].

Therefore, predicting power demand employing different technical solutions, including the so-called artificial intelligence, is a burning issue related to operating power grids. It provides a real opportunity to impact electricity distribution throughout the system, limiting unnecessary consumption [8,39,40,41]. It is an actually applied way of limiting carbon dioxide emissions by power plants in the so-called small power grids—Figure 3. Another method to balance power demand is purchasing electricity from other systems—Figure 3. Small power systems (Figure 3) always include a finite, local number of electricity consumers. As shown in Figure 2, the power demand is the total power for individual consumers O₁, O₂, O₃, …, O_n is also always characterized by a well-defined (usually known) variability over a given PS operation time. Demand variations and different electric consumer start-up times are reasons for PS load changes at a given moment and, thus, power demand at a given time t [21,28,42,43].

The inclusion and possibility of determining and employing load predictions throughout the entire NPS, let alone an LPS, is always a function of response to instantaneous power demand for the power plant itself—Figure 4. This function also depends on the power plant technical parameters applied within a given PS, Figure 4. One of the key aspects in determining electricity demand in NPS and LPS is also the determined possible gradient of power demand changes over time, or the power change rate for a given time interval (response time to a change in input conditions—power demand). It is the magnitude of these changes, W, kW, or MW, at a given moment in time, as well as the determination of the minimum and maximum power within a given SE (cable cross-section, permissible voltage dips in a PS, etc.). Figure 4 illustrates the rate of power plant load (response) change depending on the electricity generation technology employed [12,20,44,45,46]. The power plant that is characterized by the highest rate of such changes in response to a given power demand can be a hydroelectric facility with an artificial or natural damming reservoir. In such a case, the adjustment rate is 100%/min. In practice, such a power plant lacks inertia to adjust power to demand. Whereas a conventional—coal-fired—power plant is characterized by very high inertia. It exhibits a possible power demand change of only 4%/min., as shown in Figure 4 [11,14,26].

2. Literature Review

All PS operated in specific countries worldwide should come with a certain flexibility and be characterized by a permissible minimum lower and maximum upper range of power demand adjustment over a rather short period of time [1,5,8,44]. This adjustment, or in other words, change in power demand at a given time t, is implemented within a PS always by operated power plants only. Power plants provide electricity to a PS. It can then be distributed (forwarded) to different recipients, usually consumers of varying rated capacities [6,11,20]. The power demand within a given PS can also be balanced using other available external grids, usually from neighboring countries [5,14,18,45]. The PS time of response to changing power load parameters depends on different power plant types. This enables the successful employment of various power plant types within a national power system. The time of response (time constant or inertia) to demanded power should be as short as possible. This enables covering (balancing) the power demand at a given moment in time [3,13,15,46]. The authors of the paper have encountered different proposals for technical solutions to this issue in the case of large PSs. However, the available source literature still lacks solutions for local, small-area, so-called island PSs. The authors suggest employing neural networks in a power system to predict PS demand changes and later apply these results to automatically adjust power plant capacity [8,21,28,47]. Based on the so-called artificial intelligence, it is a novel approach employed to assess load changes. It is applicable to determine power demand in small-area PSs [16,26,27,41,45].

Another important issue that has to be implemented in the PS load prediction process is the so-called continuous diagnostics process when such technical facilities are online [12,45,47]. Diagnosing technical facilities of such class is also associated with the time of response to unfitness states within a PS. An unfitness may be partial (local), e.g., failure of a single transmission line, or total (catastrophic) [17,37,41]. The latter should not occur within a PS of a given country [5,12,21,48]. In such a case, it is, e.g., a power plant’s failure that is dominant within a given PS due to its maximum output capacity [38,43,46]. It is a particularly important issue, since power facilities are always classified as SCIFs [27,33,45,47]. The process of acquiring information on the current load should be implemented in real-time, and diagnostic information should be forwarded to a PS National Dispatch [49,50,51].

The reliability of electricity supply to all consumers is also important in terms of operation. This issue must be taken into account at the engineering stage, and all national PSs must be retrofitted [27,31,33,42,52]. Currently, virtually all devices and consumers utilize electricity drawn from a PS to operate [10,25,35,40]. An unfitness within a PS (partial or total) entails tangible economic losses (e.g., stopped manufacturing process, railway or air traffic), as well as image-related losses [3,33,53,54]. Therefore, all available methods and technical solutions must be employed to improve power supply reliability within a power system, e.g., redundancy or the fail-safe principle [9,28,33,40,42].

Another extremely important parameter that characterizes all national PSs is the so-called voltage or current quality found in a given consumer’s receiving node [8,14,17,55]. This involves, e.g., content of harmonics within a PS, dips, drops, permissible grid frequency and rated voltage changes, and value of interruptions (short- and long-term) related to industrial power supply [6,9,15,44,48].

All devices and components of a PS must also exhibit sufficient resistance, strength, and susceptibility to natural and artificial interference appearing within the natural environment [44,52,56]. These include, e.g., atmospheric discharge or electromagnetic radiation sources over the entire frequency range [27,31,40,56,57]. Individual devices operating within a PS must also be fitted with anti-destruction and anti-failure elements, responding to, e.g., power grid over-voltages—e.g., varistors or fuses [17,31,42,48,58,59].

The research aim of this work is to develop a neural model using an LSTM network to forecast power demand in a small power system.

3. Power System Power Demand Changes

The local response of the system to power demand changes when this leads to a requirement to adjust the voltage at the terminals of the transformers located in switching stations (marked A and B in Figure 2) is implemented within the system at the GPZ (transformer/switching substation) level. The basic method is to employ the I&C of the AVR—Automatic Voltage Regulation. It is the I&C built into HV/HV (A—Figure 2) and HV/MV, MV/MV (B—Figure 2) transformers, implemented through changing transformer taps. The time of response to a deviation from a preset voltage level is, e.g., 120 s for the lowest difference. It is, of course, correspondingly shorter when the difference between measured and preset voltage is, respectively, higher.

A frequency drop is possible in the event of a sudden deficit in power supplied relative to the system power demand. It is a dangerous situation because it may lead to emergency generator shutdown through frequency protectors and to a cascade intensification of system failure, even resulting in a PS blackout (catastrophic failure).

The basic protection against such situations is the Automatic-Spontaneous Load Shedding (ASLS). The response time arising from the guidelines of the Power Transmission and Distribution Association (PTPiREE) is immediate. A minor delay in the case of not-yet-modernized GPZs is permissible. The ASLS I&C operation involves shutting down consumer groups (usually several MV lines) to relieve the power system. The ASLS usually has only two consumer shutdown stages. These are two different groups of all shut-down consumers—consumers within a given power grid. The second stage is implemented if the first-stage shutdown does not lead to standardizing frequency. Due to the current multitude of distributed cycle generation (consumer shutdown implementation) on MV lines, the ASLS I&C issue requires analysis and adjustments. Such actions can often be harmful due to shutting down generators coupled in an MV line string.

If there is too much generated power relative to the demand, the increased grid frequency should lead to adjustments as shown earlier. In turn, sudden changes should lead to emergency generation shutdown by over-frequency protections.

The power system in Poland is always monitored by the KDM (National Power Dispatch). Its subordinate units are

• ODM—Area Power Dispatch,

  ○

  CDM—Central Power Dispatch (central in the sense of a local distribution grid operator over a given area of the country),

  ▪

  RDM—Regional Power Dispatch.

Energy generator forecasts in terms of the electricity planned for introduction to the power system the following day are collected on a daily basis. In addition, these entities analyze weather information and its impact on generation, also based on power source, for which there is no requirement to file reports and forecasts. The volume of electricity to be drawn from the system is also determined based on historical data from previous years. KDM also takes into account the power demand in neighboring countries, coupled to common power systems. All of the information above results in a decision on the volume of electricity required from stable energy sources (e.g., coal-fired, gas-fired, and hydropower plants) and the possibility of accepting electricity from non-stable sources (wind and solar power plants). KDM forwards guidelines on the range of required electricity generation or its limitation to its subordinate units [60,61]. The information on the need to introduce restrictions is continuously monitored and adjusted, and the decisions on, e.g., shutting down/limiting power from specific sources are forwarded as late as possible, but within a time that enables their implementation without detriment to the power system and the equipment hooked up to an industrial power grid. The usual time for a PS is 1–2 h in advance.

3.1. Local Power System Power Demand Analysis

The analysis focuses on a local power system that covers the area of a large provincial city in Poland, together with its agglomeration. The power demand data related to this system was collected at a one-hour interval and covered two years. This amounts to a total of 17,520 h. The power demand time waveform is illustrated in Figure 5.

A small power system is characterized by large power demand variability over time, as depicted in Figure 5. It is caused by its sensitivity to power supply shutdown by individual, relatively large industrial plants located within such a system and having high-capacity consumers. The fact that such sudden and significant power demand changes appear within such an isolated power grid makes forecasting related to such systems much more difficult than in the case of an entire, large system that usually covers the area of an entire country.

The most important statistical parameters of power demand within the studied area have been listed in

Table 1. Local power region power demand statistical parameter analysis.

Parameter	[MW]
Minimum value	119.99
Maximum value	609.52
Mean	334.25
Median	328.35
Standard deviation	93.47
Variation range	489.53

. The range of power demand changes is very large (489.53 MW) and significantly exceeds the mean value (334.25 MW) by more than 46%. The standard deviation for these changes in the power demand was 93.47 MW, which constitutes approximately 28% of the mean value. This proves large demand variations (changes) within the studied time interval and always increases the difficulty of forecasting—predicting load changes within the t time interval in question.

Figure 6 depicts a colored power demand map for a local power demand region over two successive months of a given year (January and February), broken down into consecutive hours of the day—on the horizontal axis.

Green marks low power demand—in the order of 300–350 MW, while yellow marks high demand—in the order of 500–550 MW (scale shown on the OY axis, to the right of the image).

3.2. Power Demand Forecasting vs. The Issue of Reducing Carbon Dioxide Atmospheric Emissions

The issue of forecasting power demand is an element of planning an entire national power system (particularly so-called own, without the help of neighboring countries, grid power balancing) that is important from technical, economic, and environmental perspectives. The stability of a given power system ensures an appropriate balance between demanded electricity and electricity supplied to a given system. When applying an accurate forecast, system operators are able to adapt the level of currently generated electricity to the current demand [62]. The solution to the problem of predicting a time series, namely, power demand within a given power region, is based on various methods. In light of the growing demand for electricity and its increasing prices, forecasting accuracy is also becoming increasingly severe in virtually all countries of the world. Over the recent years, researchers have focused their studies in this field on artificial intelligence methods, particularly neural networks (such as, e.g., MLP—Multilayer Perceptron, RBF—Radial Basis Function, and SVM—Support Vector Machine [63,64]), as well as fuzzy systems [63,64,65], evolutionary algorithms, and auto-encoders [66,67]. In addition, linear ARMAX models [66,68] are also employed for stochastic time series forecasting.

Deep neural networks, such as CNN—Convolutional Neural Networks—or LSTM—Long Short-Term Memory networks [68]—are growing in importance in relation to the power demand forecasting task. An LSTM network is structured in several layers: the input signal layer, hidden neuron feedback layer—represented by a so-called cell (Figure 7)—and an output layer. The output layer is always fed by output signals from the hidden layer [60,61].

In Figure 7, the forget gate has been marked by a red frame, the input gate is marked by a yellow frame, and the output gate by a green frame. The signal flow directions have been marked by arrows. In the previous version of the drawing, not all signal flow directions were marked. In the input gate, the cell updates in the previous and present time instants have been indicated by (č_t−1) and (č_t), respectively.

An LSTM cell (Figure 7) processes signals over two consecutive time instants: previous (t – 1) and current (t). The x_t symbol marks the input signal vector at time t. Therefore, it represents the current value of the forecast time series. The c_t−1 and c_t symbols mean a stored cell status at the previous and current moments, respectively. Similarly, h_t−1 and h_t mean hidden layer output signals at the previous and current moments, respectively. These signals are then forwarded to the output layer. The cell conducts mathematical operations as illustrated in Figure 7. They can be expressed as the L function of three input variables (h_t−1, c_t−1, x_t) for a previous moment and two output variables (h_t, c_t) for a current moment, as shown through the mathematical expression No. (1).

$$\left(h_t,c_t\right)=L\left(h_{t-1},c_{t-1},{\mathbf{x}}_{\mathrm{t}}\right)$$

(1)

Weight vectors are subject to adaptation within a network learning process. Learning is implemented under supervision, e.g., using an adaptive moment estimation algorithm (ADAM). After passing through the hidden layer, input data goes to the output layer. The difference between an actual and expected output value is always treated as an objective function subject to minimization. This function is then reverse-propagated by the network to calculate the gradient in the neural network weight selection optimization procedure [61].

The LSTM network was employed to predict power demand for a one-hour advance period. The number of hidden neurons was selected based on preliminary experiments in the range of 300 to 650. The best results were obtained for 465 hidden neurons. Numerical studies were implemented in the MATLAB R2023b environment. Prediction accuracy was expressed via commonly used accuracy measures, such as MAE (Mean Absolute Error)—expression No. (2),

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{n}}\left(\sum _{h=1}^n\left|T\left(h\right)-P\left(h\right)\right|\right)$$

(2)

RMSE (Root Mean Squared Error)—expression No. (3),

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{n}}\sum _{h=1}^n{\left(T\left(h\right)-P\left(h\right)\right)}^2}$$

(3)

MAPE (Mean Absolute Percentage Error)—expression No. (4),

$$\mathrm{M}\mathrm{A}\mathrm{P}\mathrm{E}={\displaystyle \frac{1}{n}}\sum _{h=1}^n{\displaystyle \frac{\left|T\left(h\right)-P\left(h\right)\right|}{T\left(h\right)}}\cdot 100\mathrm{\%}$$

(4)

and Pearson’s linear correlation coefficient R—expression No. (5),

$$R={\displaystyle \frac{\mathrm{c}\mathrm{o}\mathrm{v}\left(T\left(h\right),P\left(h\right)\right)}{{\sigma }_T{\sigma }_P}}$$

(5)

In the case of mathematical expressions No. 2, 3, 4, and 5, T(h) means an actual power demand at time h, P(h) means the demand forecast, while σ is the standard deviation. The conducted studies resulted in the following values of forecasting accuracy measures:

• MAPE = 2.7721%,
• MAE = 8.3009 MW,
• RMSE = 12.5277 MW,
• R = 0.9834.

The accuracy indices above were calculated for the entire tested range, as shown in Figure 6.

Figure 8 illustrates actual and predicted power demand for a local power region for test data not involved in learning and a period of 2 months. A good match of both curves is evident. This proves the high accuracy of the predictive method employed. Figure 8b shows marked prediction errors for consecutive power demand estimation hours.

Input data length is at least 2 years, which amounts to a total of 17,520 h. The data has been collected at a 1 h time interval. The information on input data length is already included in Section 3.1 of the paper. Local power system power demand analysis. Approximately 80% of the input data set was allocated to learning, and the remaining 20% to testing.

The forecasting model structure is illustrated in Figure 9 (below). Time series graphs can be found on the left: for one day, for one week, and for one month (measurement). The next block is input data analysis that includes data standardization as per the formula:

$$x_j={\displaystyle \frac{r_j-m}{\sigma }}$$

where: x_j—standardized value of demand at time j; r_j—true demand value prior to standardization at time j; m—mean demand value within the studied time interval; σ—standard deviation within the studied time interval.

Next, the authors presented a diagram of the employed LSTM network (signal processing). The sequence Input Layer prepares data for forecasting. It involves determining the number of features subject to forecasting. It is followed by the IstmLayer, where the number of hidden neurons is specified. Next comes the dropout layer, which is responsible for regularization with a preset probability. The fully connected layer multiplies inputs through a weight matrix and adds a bias vector to it. The final forecast is the outcome of a forecasting team comprising 5 trained LSTM networks. The operations of the LSTM network are repeated 5 times, and the end result (final forecast) is an arithmetic mean for the entire team—

Table 2. Parameters of the proposed forecasting model—search space for the optimal and selected parameter.

Parameters of the Neural Model	[Search Space], Selected Parameter
Number of neurons in hidden layer	[300…465…650]
Number of epochs to train model	[50…70…150]
Optimizer function	[rmsprop, adam, sgdm, lbfgs]
Learning rate	[0.001, 0.005, 0.01]
Loss functions which are minimized during training	[mae, mape, mse]
Number of ensemble members	[3…5…10]

. The application of an artificial neural network to forecast electricity demand in small power systems enables automatic learning based on current change observations. This also includes automatic generalization of knowledge gained throughout the entire process of using such software in the power utilities’ local power distribution center. These processes are often not properly implemented, particularly in small power regions.

Figure 10 depicts a graph that illustrates the amount of carbon dioxide (CO₂) emitted to the atmosphere per one-megawatt hour (MWh) of electricity generated by a power plant, taking into account the fuel used by conventional power plants of a given type.

The highest CO₂ emission per 1 MWh of electricity generated is recorded in the case of lignite coal and amounts to 1065 kg CO₂/MWh. In turn, this parameter in hard coal-fired power plants is 900 kg CO₂/MWh, and in the case of natural gas-fired power plants, 450 kg CO₂/MWh.

According to the data in [54,66,69], 46.9% of the electricity in Poland was generated in hard coal-fired power plants, 22.5%—in lignite-fired power plants, and 13.2%—in natural gas power plants. Therefore, the three fuel types referred to account for 82.6% of electricity generation in Poland (as of 10 February 2025). Figure 11 illustrates the impact of an accurate forecast on the reduction in the emissions of a selected greenhouse gas (CO₂) upon the application of a coal-fired power plant power demand prediction developed by an artificial neural network. Prediction enables electricity generation to adapt to the current demand. This leads to savings, both in terms of generating the electricity itself and losses during transmission over large distances (power plant—transformer station—switching stations—consumer—Figure 2), and therefore, in the fuel consumed to generate 1 kWh in a power system.

This also means a significant reduction in the production and emission of greenhouse gases into the atmosphere. Limiting the level of generated power by, e.g., 54 MW in one hour leads to reduced CO₂ emissions by as much as 48,600 kg in a conventional, hard coal-fired power plant. In a lignite-fired power plant, where unit emissions were even higher, this value is 57,510 kg CO₂, and 24,300 kg CO₂ for a gas-fired power plant. Figure 11 illustrates the impact of employing power demand prediction on reduced carbon dioxide atmospheric emissions in the case of coal-fired power plants in a PS.

Figure 11 is aimed at demonstrating how information originating from a load forecast for the following hour (during a morning peak) may impact reducing atmospheric CO₂ emissions of a coal-fired power plant. Reducing generated power by, e.g., 54 MW, possibly owing to next-hour demand information, enables reducing CO₂ emissions by 48,600 kg. Figure 11 is aimed at illustrating an example of a situation wherein knowing power demand in the following hours, during a morning peak, enables reducing the atmospheric emissions of CO₂ generated by a coal-fired power plant.

The developed method triggers industrial energy storage as well as hydropower plants located on waterways in northern Poland. Industrial energy storage is often installed in plants with extensive high-capacity photovoltaic systems. The aforementioned adjustments (increased power demand) are implemented in practice without time delay. The introduction of an additional power supply to a power system is immediately visible on power demand graphs received in real time. Of course, employing other technical solutions to increase power in a system—e.g.; increasing the capacity of a CHP with an additional power generator—requires the existence of feedback—i.e.; information on the system’s response to increased power. The authors are currently implementing such research in relation to another power system, thus creating an opportunity for comparison—e.g.; control quality (e.g., power system’s response time to forced actions—power increase). Companies (electricity distributors) within a given area do not provide information since the domestic power market is currently experiencing fierce competition.

4. Conclusions

The prediction and analysis of power demand for small power grids conducted using artificial intelligence (using a deep neural network, LSTM, in this case) clearly confirmed that this might become one of the ways to limit carbon dioxide atmospheric emissions. An analysis of the power demand issue and the research that was conducted indicated that such an application of a neural network is a good tool for forecasting short-term demand in a local power system. Knowledge of the power demanded within a power system over the next hour allows an operator to increase or reduce electricity generation, depending on the needs of the entire power system. This ensures reliable operation and balancing of the entire power system—in this case; for small power grids. Forecasting power demand in a local power system significantly contributes to reducing greenhouse gas atmospheric emissions. Employing accurate forecasts is important in small power systems, which experience relatively large changes (sudden surges) in the power demand throughout the day. Conventional power plants (that use hard coal, lignite, or natural gas as fuel), where the rate of electricity generation changes the least (around 4%/min.), require knowledge of forecast demand well in advance—significantly faster than; e.g., hydro or pumped hydropower plants. This is caused by the fact that they are unable to reduce generated power over a rather short period of time and, hence, cut the atmospheric emissions of harmful greenhouse gases, which are subject to specific emission limits in individual EU countries. Reducing electricity demand within an industrial power system by, e.g., as little as 54 MW in just one hour of its operation leads to CO₂ emissions reduced by as much as 48,600 kg in a conventional power plant with hard coal as its fuel. It is different in the case of a power plant with lignite or natural gas as its fuel.

In the former case, a lignite-fired power plant, unit CO₂ emissions will be many times higher. This value amounts to 57,510 kg CO₂. And respectively, 24,300 kg CO₂ for a gas-fired power plant.

Figure 10 is a presentation of the authors’ analysis of the impact of employing power demand prediction in small power grids on reduced carbon dioxide atmospheric emissions in the case of coal-fired power plants in a PS.

The authors of the article employed a neural network for computer predictions of load within a power grid, without previous mathematical formalization of the state of the issue. Using a neural network to predict load changes in power grids also makes it possible not to reference any theoretical assumptions in relation to the issue being solved. The availability of power grid load results, particularly their analysis in individual power plants supplying specific regions, is low. The application of a neural network and its capability to learn based on a current change in the power load enables automatic generalization of the knowledge gained throughout the entire process of using such software within a power distribution center of power plants. In their preliminary research, the authors of the article have referred to only one country, i.e., Poland. They based their research on an ‘island’ power system located north of the country, which experiences significant distances to the nearest power plants. A crucial research problem when conducting this task was also obtaining power demand data. The approach to this issue of companies supplying electricity to consumers in this area differs from company to company. Rather, they are competitors and do not always intend to share the data they consider confidential. However, while continuing to work on this problem, the authors are trying to interest the nearest countries and neighbors in this research topic.