2.1. Methodology Overview
Day-ahead markets represent the most active markets in terms of economic value of transactions, but other markets have experienced a noticeable growth (for example Intraday Markets in France and Belgium with growth rates of 54.5% and 82.9%, respectively, in 2018). Real-Time Markets have facilitated the integration of renewables in USA markets in the last decade, and wider-scale markets with later gate closure would facilitate the integration of renewable in other systems (e.g., Balancing Markets, and specifically, Reserve Replacement). The integration of demand-side resources in markets presents both risks and opportunities for Balance Responsible Parties (BRP, responsible for its imbalances) and Balance Service Providers (BSP, i.e., the provision of bids for balancing) and need the development of new and more integrated methodologies. The main idea of this research work is providing new tools to aggregators for a best management of demand and generation in markets, both in the short-term (around 24 h) and in the very short-term (from several minutes to 1 h), to evaluate net demand unbalance, while the aggregator or other parties take into account gate closure times. To perform this task, the proposed methodology takes profit from different databases which should be able for DR (demand, customer, RES, and weather). From these databases, this work applies different methodologies to obtain well fitted 24 h forecasts for demand and PV generation, and with these predictions aggregators evaluate bids and offers to be sent to Intra-Day or Real Time Energy Markets (with the objective of minimizing the energy costs for customers or prosumers). Logically, both models (demand and PV) exhibits errors and these errors can involve penalties in the markets because other agents (BSP, LSE) should change their energy balance and buy or sell energy in the very short term. Considering that PV-forecasts usually have a greater error than demand-forecast, and that a fast model is needed to manage the potential flexibility of demand in the aggregator-side, a simple and very short-term PV forecast model is developed based on historical recordings of PV generation (in the site) and available real-time measurements (Information and Communication Technology, i.e., ICT devices). This model and the results of cassation processes in short-term energy markets provide a reference signal for flexible resources (only DR resources are considered in this paper). With the help of different tools and scripts from a DR-toolbox (e.g., segmentation, classification, disaggregation and modeling), the aggregator can evaluate the “demand baseline” for different end-uses in the short-term and can propose a control signal to change demand according to its requirements. This demand is simulated and evaluated hour by hour with several indices of performance (and modified in some cases) to fit the demand packages offered to energy markets (i.e., net demand). In other cases, when the power system tackles for flexibility, the aggregator can provide additional flexibility to energy and ancillary services markets, agents or Transmission Operators.
Figure 1 presents an overview flowchart, which depicts the methodologies and tools to be used through the paper.
A quantitative analysis for demand-side flexibility seems necessary thorough the definition and the evaluation of some indicators (i.e., DR indices in
Figure 1) that allows to score the flexibility and performance of loads being controlled, basically at the aggregated level. These indicators converge with the idea of some voluntary schemes in the EU that intend to express the “readiness of a building” (in this case the readiness of the load inside buildings). According to these proposals [
31], these indicators mean: “readiness to adapt in response to the needs of the occupants, readiness to facilitate maintenance and efficient operation, and readiness to adapt in response to the situation of the energy grid”. Taking into account this last requirement, a score is performed through the indicators to be described in
Section 3.4.3.
2.2. Characteristics of the Customers: Demand, Photovoltaic Generation, and End-Uses
All electricity customers from a small town (4400 inhabitants), sited in the north Spain, have been selected for simulation purposes in this work. These customers include residential, commercial, and industrial clients, although most of the power consumption is due to the residential ones. Basically, this group corresponds to average residential customers in the south countries of the EU. The rated power per customer ranges from 3 to 8 kW. The climate is continental, and winter temperatures range from 0 to 13 °C and in summer from 13 to 29 °C.
Regional investors built a PV plant in the vicinity of the town, with a significant capacity with respect to its power demand. The PV plant is composed of two-axis solar trackers with a rated power of 1.9 MW, and it is connected to the same power substation that links the town to the grid.
Hourly load and photovoltaic generation data from 1 October 2008 to 31 March 2011 (both included) were available. These data were obtained from the electric utility distributor and corresponded to hourly average power measurements in the substation. It is worth mentioning that it has been very difficult to obtain real data corresponding to a considerable customer group that can act as prosumers (consumers and producers); thus, we had to manage data not as recent as desired.
Figure 2a shows the winter and summer loads for two selected workdays monitored at the distribution level (substation) that supplies power at 13.2 kV to the distribution transformer centers (CT) of customers (basically residential and commercial supply).
Figure 2b shows the average temperature in the area for the same two selected workdays.
Figure 2c plots the hourly PV power generation on the peak production days (days with the highest energy generation) of January and June 2010. The PV power generation values in the central hours of the day can mean an important fraction of the town consumption (30–40%). The selection of January and June as representative months is due to the fact that June corresponds to the month with the highest PV generation, while January corresponds to the lowest PV generation and the highest energy consumption. In
Figure 2, it is also shown the average profiles of demand, temperature and PV generation for the period in which data is available (from 1 October 2008 to 31 March 2011). It is remarkable that the average PV generation profile (
Figure 2c) is lower compared to the other ones (June and January). This fact can be explained because months’ profiles are representing the peak power days, while the average profile includes days in which there are no PV generation due to adverse weather conditions.
The use of DR portfolio for damping both the errors in the evaluation of demand in short-term and the intrinsic variability of PV generation sources need the evaluation of DR potential and its flexibility in each customer segment. First, this evaluation must be based on the knowledge of end-uses for an average customer. The first alternative to know demand composition behind-the-meter is the use of the information provided by Smart Meters (SM) and then apply some Non-Intrusive Load Monitoring (NILM) methodology, for example [
32]. This last approach involves the full development of capacities of available home automation technologies, considering the increasing deployment of Smart Meter in several countries around the world [
33]. Some of these methodologies have been reported by authors in previous papers [
34] to obtain end-use disaggregation/profiles in residential segments and their real flexibility when DR policies are applied (i.e., DR validation).
In some cases, and from a practical point of view, it is possible that some problems arise for a practical implementation of DR based on end-uses, for instance: small customers do not have yet any SM, confidentiality of data is in question, Data Exchange Platforms (DEO) are not fully developed, and the availability of data is scarce [
35] or the aggregator has access to meter data but without the necessary granularity or quality (i.e., it is usual to have data with granularity ranging from 15 min to 1 h which usually makes much more complex the identification of loads through NILM methods). In this way, an alternative access to demand data should also be considered by aggregators to accomplish the evaluation of DR potential. This alternative is based on periodic surveys performed by international or national Energy Agencies, for example EIA (data from 2015, [
36]) in the United States or the Joint Research Centre (data from 2016, [
37]) in the European Union. In this way, a residential “average” EU or USA customer (and its end-use share) can be estimated according to these data. In the case under study, available reports from the Institute for Energy Diversification and Energy Savings (IDAE, Spain) and the Spanish Government [
38] have been analyzed.
Table 1 depicts the main end-uses for Spain, EU-28 countries and the USA. Notice that in European Mediterranean countries, the Air Conditioning load represents a higher percent (66% of households have this appliance and the increasing trend is quite solid). A similar trend can be reported in the USA, because 87% of homes use air conditioning. It accounts for 12% of annual residential energy expenditures and is a large factor in fluctuations in residential electricity use. Heat Pumps (HP) exhibit similar trends according to EIA estimations [
36]: the share of heated homes using HP increased from 8% to 12% in a decade (from 2005 to 2015). At the same time, the share of homes using electricity for water heating (WH) increased by 7% (to 46%). Due to this fact, both loads (HP and WH) have been considered to evaluate demand flexibility in this work. Moreover, winter period has been selected for simulation purposes in the following paragraphs, because demand in winter is higher than in summer and the climate in this Spanish area is more restrictive for PV generation possibilities.
To obtain some representative profiles, it seems necessary to evaluate load dynamics, and the service the customer obtains from them.
Figure 3a,b shows some real end-use load profiles for a household belonging to the overall customer demand, previously shown in
Figure 2. In this case, feedback from everyday activities [
40] of the customer is important to refine profiles, improve DR&EE (Demand Response and Energy Efficiency) success and gain customer interest in energy concerns. Regulations can help aggregators to establish load profiles.
Figure 3a shows an average HP consumption profile in winter, as in this study, DR simulations to balance generation are focused on this period.
Figure 3b shows an average water flow use to determine WH requirements extracted from EN 15316-3-1, Section 5 (EU normative).
Figure 3c shows the proposed end-use profiles for an average customer.
The procedure for obtaining end-uses profiles (
Figure 3) could be explained as follows. In the first place, the aggregator needs to recover basic information about customer daily overall demand (aggregated or not) through Smart Metering (
Figure 1, left bottom side). This information, alongside weather databases and public reports of energy household demand and share of end-uses in terms of energy, allows the aggregator the calculation of “household baselines”. At this point, the aggregator is able to run and refine Physically Based Load Models (PBLM) both at elemental and aggregated levels (i.e., include inputs/outputs for these models). With PBLM and average weather inputs the aggregator obtains “end-use baselines” for each end-use with relevant potential for DR (e.g., HVAC, space heater, WH, or thermal ceramic heaters,
Figure 1), and their average daily demand in each season/month. Finally, “elemental baselines” (kWh and profiles) are modulated through coefficients (considering weather conditions and customer behavior) to fit the “overall baseline” for the customer.
2.4. Time-Series Clustering
Clustering is an unsupervised technique whose objective is to separate objects (represented by a multivariate dataset) into homogeneous groups (called clusters), such that objects in the same cluster have high similarity among them, but low similarity with the objects in a different cluster. It is considered an exploratory technique very useful by itself or as a previous step for other kind of data analysis. Depending on the way the clusters are generated, clustering methods can be divided in two big sets: hierarchical methods and divisive methods. In addition, the resulting clusters are determined by the distance or similarity measure and the linkage method selected.
A special case is time-series clustering, where each object to be grouped corresponds to a sequence of values as a function of time. One of the main advantages of clustering time-series is that it allows the discovery of hidden patterns in time-series datasets. Generally, three different objectives can be considered in this context: finding similar time series in time, in shape, or in change (structural similarity). The selection of a suitable distance measure is essential and depends on the objective pursued. Interesting surveys in the field can be found in [
46,
47].
In this paper, we have focused on similarity in shape to cluster the daily curves of photovoltaic generation. Dynamic Time Warping (DTW) distance, described below, was chosen for that purpose. Regarding the nature of the clustering, hierarchical technique together with average linkage were selected. These clustering methods were developed by means of the R package TSclust [
48].
DTW distance, introduced by [
30], is commonly used for measuring shape-based similarity between two time series, which may vary in timing. The main advantage against other shape-based distances such as Euclidean or Wavelet Transform is its invariance to warping. In our context, daily curves of PV generation are conditioned by sunlight hours, which vary along the different seasons. That makes DTW distance suitable for clustering a set of daily PV curves along different years.
Given two time series (
xi)
I = 1, …,
m and (
yj)
j = 1, …,
m, it starts calculating a
nxm matrix
D = (
Dij) with the distance between every possible pair of point
xi and
yj in the two time series,
Dij =
d(
xi,yj),
I = 1,…,
n,
j = 1, …,
m, where
d(
xi,yj) can represents the Euclidean or the absolute distance. According to [
30], a warping path
w is a contiguous set of matrix elements which defines a mapping between (
xi) and (
yj) that satisfies the following conditions:
Boundary conditions: w1 = (1; 1) and wk = (m; n), where k is the length of the warping path.
Continuity: if wi = (a, b) then wi−1 = (a’, b’), where a − a’ ≤ 1 and b − b’ ≤ 1.
Monotonicity: if wi = (a, b) then wi−1 = (a’, b’), where a − a’ ≥ 0 and b − b’ ≥ 0.
The objective in DTW distance is to find the shortest warping path. Due to its high computational cost, different approaches have been proposed to optimize the calculation (see [
49]).
2.5. Demand Response Strategies
Demand Response policies have been used by ISOs since the early 1980s. In the first years of DR, the objective was to achieve a more rational planning and operation of resources. In recent years, with the development of energy markets and the increasing share of RES in the generation mix, DR becomes more centered in the customer and in the integration of the available RES potential. Demand Response can be divided into explicit and implicit DR. Implicit DR means the change of demand due to prices whereas explicit DR involves the change of demand when System or Distribution Operators (i.e., ISO or DSO) forecast and declare an event into the system in the short-term.
To respond to these events or prices, the most common policy is to limit demand. This reduction can be performed through the cycling of power supply (the supply is switched ON and OFF alternatively following a rectangular wave
u(
t)). If the natural “cycling” of the end-use being controlled,
m(
t) (the operating state of the control device), is greater than
u(
t), the DR action is effective (notice that an appliance can describe cycles or not, for example a fridge or an inverter heat pump, but every load has its operating state
m(
t) with respect to rated power). Considering that, in practice, demand measurements are discrete (every 5, 15 or 60 min) and it is necessary to define average values in a time period [
t,
t +
k]. Mathematically, Equation (1):
where
(t,
t +
k) and
ū(t,
t +
k) are mean values of
m(
t) and
u(t) in the time period k, respectively, and
tON is the time in this period where a “representative” (average) load remains switched
ON and demands power.
The models to be used (to obtain
m(
t) and apply
u(t)) are PBLM, a methodology proposed first to solve problems such as cold load pickup. The main reason for this choice is that these models are “white” [
50] or “grey” [
51] models which allow physically explaining the dynamics of the appliance and its environment and consequently foreseeing its changes. In this work, PBLM “grey” models for HVAC (Heating, Ventilation and Air Conditioning) and WH loads (heating and ventilation) previously proposed in [
39] have been used.
Figure 4 shows an electrical-thermal equivalent for this model for heating loads (a broader explanation of parameters can be found in several references [
52,
53]).
The main features of these models are: they consider heat gains and losses, for instance solar radiation (H
sw, H
w) or internal gains due to inhabitants (H
r) or appliances (H
a) (
Figure 4a); the model takes into account heat storage from the specific heat of external walls (C
w), indoor masses (C
a, C
1 and C
2, especially important for WH) or roof/ground (C
rg); and it considers the control mechanisms which drive appliances (for instance thermostats
m(
t) and DR policies
u(t)). Moreover, their state variables are temperatures: indoor (X
i), walls (X
w), roof/ground (X
rg) for HVAC loads (
Figure 4a), and X
1, X
2 for the stratification of water in the reservoir (“hot” WH1 and “cold” WH2 sub-tanks,
Figure 4b), that is to say, characteristics that allow the evaluation of energy flows and storage capabilities (i.e., the indirect capacity of storage in the envelope of buildings), the direct storage in WH or the loss of customer service due to the application of DR (i.e., internal or hot water temperature).
These models are individual ones and need a further aggregation to reach a minimum demand level (size of reduction packages) established by specific regulations of electricity markets to bid or offer into these markets (e.g., from 100 kW to 1 MW depending on each specific service or market [
54,
55]). This task is often developed by energy aggregators.
To achieve these packages, the aggregator needs to rise ON-time to increase demand of each specific end-load whereas a decrease of demand requires a reduction of ON-time. The second alternative (the traditional one) is easier because the aggregator only needs to manage the rate of switching-OFF and switching-ON times of the power supply to load. This is easy to perform through hardware by classical methods (e.g., ripple control of WH in Germany, [
56,
57]) or home automation methods (e.g., controlled plugs through Z-Wave protocols [
58] and universal software platforms for control [
59]).
An important concern for the practical implementation of modern DR policies is the Automated Demand Response (ADR), because customer manual control does not fit the requirements both of accuracy and reliability of response. It is imperative for the success of ADR the development of standards for the communications between operators, aggregators and their customers’ automation equipment [
60] and the feedback from them. Open automated demand response protocol (OpenADR [
61]) represents a good example of such a standard. Every day, more and more devices are certified to use OpenADR 2.0 protocols, and especially Smart Thermostats, but this certification is not necessary if some gateway assumes the role of “last-mile” controller and is compliant to receive and transmit OpenADR protocols. For example, home automation platforms such as Universal Devices ISY994 Series [
62] allows the communication of residential customers with OpenADR, sending consigns and commands to home automation technologies working with different protocols (Zwave+, Insteon/X10, Zigbee Pro, Amazon Echo or Google Home). Other platforms, from well-known manufacturers, such as ABB SACE’s Emax 2 Power Controller, develop similar functions [
63] but at building scale. Examples of ADR capabilities of grid-integrated buildings and building microgrids, architecture, and standards can be found in [
60].
The rate of change is defined to the PBLM software by PWM waves: the carrier wave being tried has a frequency of 0.833 mHz (i.e., 1 cycle every 20 min) and the modulating waveform is the desired decrease in the average value in
m(
t,
t + 20 min) according to deviations between 24-h PV forecasts (see
Section 3.2) and 1-h PV forecasts used in markets (
Figure 1).
The reasons for choosing 20 min have physical and technical senses. The first can be explained from the point of view of load service in the case of HVAC: if a harsh control is needed, switch-OFF times greater than 20 min can cause thermal fluctuations in the dwellings, easily noticeable by consumers (this can produce a lack of customer engagement in DR). The second reason is the so-called “lock-out” or mechanical delay of heating and air conditioning units. This mechanism is used to prevent a rapid recycling of the compressor avoiding mechanical damages. From the point of view of DR, it can cause an additional delay when applying ON/OFF and thermostat control signals. To evaluate the effect and characterize (from a statistical point of view) this process, some residential HP appliances (rated power from 1 kW to 3 kW) were monitored by authors. Changes in customer demand due to control actions have been recorded by an electronic meter and several Z-wave wall plug switches which send data to PCs using an USB gateway. Results depict that ON latency time ranges from 20–60 s and OFF latency times range from 10–40 s [
64]. That is to say, the minimum ON-time should be in the range of one minute to limit inherent errors due to latency.
The first alternative (i.e., the increase of demand) is a less traditional option for DR [
65]. Several reasons explain both the lack of use of these alternatives and their real interest. For instance, the increase of demand requires the control of thermostats. This control is more expensive than the supply control because smart thermostats are expensive. The cheapest option (e.g., Z-Wave) cost around € 150–200, whereas a remote switch costs around € 40–60. Fortunately, modern appliances include control of temperature though mobile-phones or PC, and these alternatives can be used for control (notice that some of them are compliant with well-known DR standard protocols [
61]). In other cases, where the control device is intrusive (this is the case of WH), the cost of thermostat is similar, but the same maintenance (labor cost) is needed to include this option in the appliance. Over the last few years, some HPWH manufacturers in the USA have included these options for large units (200–500 L reservoir/storage tanks), for example [
66].
The control of the thermostat (up or down, according to season, and usually used for pre-heating or pre-cooling policies in the dwelling being conditioned) has been proposed as a “virtual-storage” resource [
67] for customers to take profit from Time of Use (ToU) tariffs or to “prepare” loads to face to DR events policies and maintain customer service (i.e., internal temperatures of houses or dwellings).
Usually, these policies have been used before the load is controlled, but the proposal in this paper is to use them continuously to adapt demand to changes in the forecasted PV generation. The control of the thermostat is evaluated and changed, if necessary, ±0.5 °C every 20 min. The reason for selecting this value is that 0.5 °C is a usual value for the change of temperature settings in home smart-thermostats.
In this way, the proposed control strategy
u(
t,
t +
k) for heating is done by Equation (2):
where
Xsup is the upper temperature of load’s thermostat, which is set as a simple hysteresis cycle with dead-band
db (usually ranging from 0.01–0.03 pu), and
Xlim is the maximum reasonable temperature inside the dwelling (for example 22–23 °C in the case of HVAC, in winter) or the maximum temperature of water inside the tank (68 °C), which avoids risk of burns if a mixing valve is not used for the control of hot water pipeline.
Xiserv is a minimum service level for the appliance (a minimum comfort temperature inside the dwelling, for example 16 °C, or a minimum temperature of hot water inside the WH, for example 36 °C).
Basically, Equation (2) means that load control is done by a double control. In the case of heating (electric heaters or HP) if the load must go up, the thermostat goes up until it reaches the maximum allowable value (
Xlim). Otherwise, if demand must fall to balance a decrease in PV generation (with respect to 24 h forecast), the thermostat or the supply is controlled to reduce demand. Notice that a “baseline”, (i.e., load demand evaluated without control
m(
t,
t +
k)) is also needed as reference for controlled load. This baseline also comes from PBLM models (
Figure 1).