Flexibility, in the electricity domain, is often defined as the ability to modify generation and/or consumption patterns, or demand, in response to an external signal to contribute to maintaining the stability of the power system cost-effectively [7]. This is exemplified by changing the amount and timing of energy use, as shown in Fig. 1, aiming to make the system more reliable, defer investments, reduce emissions, and/or reduce operational costs. Fig. 1 illustrates how the peak of demand in the morning and early afternoon can be reduced by activating flexibility, that is, incentivizing the consumer to move part of the demand to valley hours, resulting in a flatter curve and consequently in a more efficient system.

Technically, flexibility is defined as a power change activated at a specified time for a specific duration at a particular node of the grid within the power distribution system. Thus, flexibility can be defined by five attributes [8]: a)

Direction of flexibility (e.g., up or down).

Activation of flexibility can serve multiple purposes, and the goals together with the actors involved in the flexibility trading define the different markets and their operational rules. Thus, demandant of flexibility, i.e., the beneficiaries, are typically the Transmission System Operators (TSOs) and the DSOs, whereas the providers, those that have the capacity to balance generation and demand, are typically the producers and consumers.

The main actors involved in flexibility markets are [9]:

Transmission System Operator (TSO): responsible for the operation of the transmission system and its stability. TSO is a flexibility-demanding actor that requires flexibility to maintain energy balance and satisfy technical restrictions to guarantee system stability.

Flexibility products are managed at different TSO-DSO levels supported by different flexibility markets [9]: a)

Balancing at the transmission grid: Flexibility is offered to TSOs to correct imbalances between demand and supply in the electricity markets.

The first two categories, in the EU, are operated at different time frames through specific platforms according to availability requirements. These are typically known as Replacement Reserves (RR), Manual Frequency Restoration Reserves with Manual Activation (mFRR), and Restoration Reserves with Automatic Activation (aFRR), according to the harmonized market model definition. On the other hand, local market mechanisms for DSOs still have too low liquidity to be competitive, and their definition and operation depend on existing initiatives. However, the increasing electrification of heating and electromobility envisions the need for rapid development and expansion across the EU. This is the focus of this paper.

A local flexibility market (LFM) can be defined as an electricity flexibility trading platform to trade flexibility in geographically limited areas such as neighborhoods, communities, towns and small cities [10]. There are some key aspects to consider to better understand the local flexibility markets [11]: a)

Pre-qualification: It is the process imposed on flexibility assets and service providers to assess the level of complexity and any potential barriers to their participation in the local flexibility market. The process is divided into technical aspects (e.g., pre-qualification tests and other processes) and compliance with the necessary regulatory, legal, and/or financial requirements.

The sequence of trading flexibility in an LFM, generally speaking, consists of three main stages [12]: (1)

Contracting and bidding process: during this process, the DSO and flexibility providers (i.e., aggregator) reach an agreement for flexibility trading through an LFM operator. The result of the agreement includes quantity, time, location, and price of flexibility to be delivered. Flexibility demand is launched by the DSO some time (months, days, hours) before its expected activation. Flexibility demand is motivated by expected critical events in the grid, computed with forecast demand and generation together with grid models. Thus, the accuracy of forecast models is a crucial factor to estimate and elaborate flexibility demands (DSO side). Once the flexibility demand is launched to the market, elaboration of offers implies scheduling load/generation reserves by aggregators (or large prosumers/consumers).

LFMs can operate at very different time frames (i.e., months in advance, day-ahead or intra-day) and they operate in parallel with the already existing energy and balancing markets [13]. Consequently, multiple forecast models, dealing with different time horizons and granularity, are required for flexibility trading.

L(t) represents the net load demand at a specific point of the grid (e.g., feeder, transformer, or secondary substation). This can be considered as an aggregation of demand of individuals, li(t), in a specific area, Z, covered by the grid Eq. (1). Monitoring the evolution of L(t) is usually required to assess the loading of a transformer and support operation, maintenance, and investment planning. (1)L(t)=∑Zli(t)

In that sense, forecasting this demand can help prevent critical events. Let us define L~t′(t) as the forecast of this demand at t, computed at t′. Usually, a forecast is not required for a single time instant but for a time interval, T, after t1 (t′<t1<t1+T), and the precision of this estimation is crucial to take the appropriate decisions. In general, the main factors affecting the accuracy of forecasts depend on the quality of the models, and this is influenced mainly by:

Representativeness of the data used when training the model. Thus, the calendar, weather, and historical energy data are usually the input of energy forecast models. However, more data does not always imply better results (forecasts of congestion events in summer probably are not correlated with winter demand).

Thus, forecasting to estimate critical events at the substation or transformer level requires using either data from a specific meter installed in the transformer or the aggregation of data from smart meters installed at the consumption points. Intuitively, it can be deduced that the longer the difference t1−t′ the greater the forecast error; and this interval basically depends on the market definition. According to previous LFMs definitions, t′ should correspond to tt and tc in the case of the DSO (flexibility demanding actor) and aggregator/prosumers (flexibility offering actor). However, this is not completely true since the availability of data either at tt or tc depends on how often metering devices report data. Therefore, two different data sources could be available to measure demand at the secondary substations: power measurements, available in near real time (seconds), through SCADAs; or hourly/quarterly energy data available daily through the Advanced Metering Infrastructure (AMI). Despite this, it is feasible to estimate energy demand from real-time power measurements. However, this is not commonly done because the level of instrumentation of secondary substations is still very limited (less than 10% of transformers at the secondary substations are instrumented). So, usually, energy metering data is used to estimate the demand, and this is affected by a low update ratio (typically once a day with hourly/quarterly granularity). This means that day-ahead forecasting can only benefit from data of the previous day (collected during today) (t1−t′>24h).

Different methods and approaches have been found in the literature, and they can be classified into three main categories: statistical models [25,26], machine learning (ML) models [27–30], and deep learning (DL) models [31–34].

Statistical models aim to discover a mathematical relationship between the electricity consumption time series. We can mention models such as the Bayesian approach, the Autoregressive Moving Average, the Kalman filter, and the Gray methods. However, these algorithms are less accurate for long-term predictions, are highly susceptible to outliers in the dataset, and cannot handle the non-linear variation of energy data [35].

With the development of information technology, AI methods have replaced statistical models due to their abilities of learning features and processing data. Consequently, numerous studies and reviews on AI forecasting techniques have been conducted in the literature [36].

In [37], a comparison between conventional models and more recent AI models is made. Conventional models encompass time-series models (derived from auto-regressive and moving average models), regression models (linear, non-linear, logistic models), and gray models. For more recent ML and DL models, they consider Artificial Neural Network (ANN), Support Vector Regression (SVR), and Random Forest (RF). The conclusion drawn is that for daily and hourly energy consumption forecasting, more recent ML and DL-based models are more accurate than conventional models. Both ML and DL (a subcategory of ML) approaches require considerable amounts of data. In general, the more data available, the better the results will be [38]. Increasing input features of the model also improves the results if the features themselves are in some ways significant for the data model [38]. Therefore, data requirements play a big role in constructing an accurate model.

Three forecasting intervals, short, medium, and long-term, are reviewed in [39]. It is observed that most of the researches are conducted on short-term instead of medium and long-term. It is concluded in the research that a precise prediction will not only lead to notable decreases in penalties but also promote performance of various application fields, e.g., power grid maintenance and operating scheduling as well.

Reference [40] reviews the current state of the art of load forecasting models and draws several conclusions. Historical load data contributes to better model design, resulting in better accuracy and model performance. Every model has its own merits and demerits; however, hybrid methods were observed to have more performance efficiency in comparison to standalone methods. The hybridization of models with fine-tuned parameters had better performance. The short-term load forecasting is considered more reliable for operations and planning of utilities. It is also considered to further expand the operations and contribute in medium-term and long-term load forecasting.

In [41], the efficiency of hybrid models is discussed. The models put forward are: models based on Support Vector Machine (SVM) or Artificial Neural Networks (ANN) combined with Firefly Algorithm, Clustering Techniques, or Wavelet Transform, for instance.

An ML-prediction oriented review [42] exposes the superiority of eXtreme Gradient Boosting (XGB) methods compared to Multiple Linear Regression (MLR), ELastic Net (ELN), Random Forest (RF), Gradient Boosting Machines (GBM), or Support Vector Regression (SVR). In addition, the performance and accuracy of Deep Neural Network (DNN) models are praised in this research work.

DL has proven its ability to forecast time series. In particular, these models can be used to highlight inherent abstract characteristics and invariant structures in the data. Reference [43] provides a review of different DL methods such as Auto-encoder, DNN, Convolution Neural Network (CNN), and Recurrent Neural Network (RNN). The article highlights the interest of these models for feature extraction in demand forecasting, achieving good results with CNN for short-term forecasting.

Study [44] aims to optimize the performance of CNN, a model widely used in time-series forecasting. The conclusion drawn is that the smoothed CNN (a combination of exponential smoothing with CNN) outperforms other tested methods.

Other methods studied in the literature include Long Short-Term Memory (LSTM) methods. Reference [45] demonstrates the potential of using LSTM, attention-based LSTM, and Seq2Seq LSTM for multi-step prediction.

More recently, with the emergence of the attention mechanism and its integration into transformer models, the range of methods used to predict electricity consumption has been extended [46]. Initially used for natural language processing tasks, transformers are increasingly used for modeling time series. For instance, the Lag-Llama project [47], has been developed for non-limited prediction to a dataset. Lag-Llama is a foundational model built for univariate probabilistic time-series forecasting based on a decoder-only transformer architecture. It demonstrates strong results for 24-h prediction on unseen datasets and is able to achieve similar results compared to dataset-specific models.

Another hybrid model was developed in [48]. They propose a Self-Partitioning Local Neuro-Fuzzy (SPLNF) model based on Linear Neuro-Fuzzy models for short-term load forecasting, achieving better results than other tested methods, such as: seasonal auto-regressive integrated moving average (sARIMA), SVR, RF, and ANN.

This work requires a comprehensive performance evaluation for Short-Term Load Forecasting (STLF). The appraisal involves the application of multiple evaluation metrics. MAE (Mean Absolute Error) is a measure of errors between paired observations expressing the same phenomenon (Eq. (2)). (2)MAE=∑t=1T|At−Ft|

RMSE (Root Mean Square Error) is a commonly used metric for regression tasks since it penalizes large errors more heavily, making it sensitive to outliers (Eq. (3)). (3)RMSE=1T∑t=1T(At−Ft)2

The sMAPE (symmetric Mean Absolute Percentage Error) (Eq. (4)) addresses some of the limitations of MAPE by applying a higher penalty to positive errors than to negative errors [49]. However, if the actual value is zero, the forecast is likely also close to zero. In such cases, the denominator in the formula approaches zero, which can lead to instability, so these instances are typically excluded from the calculation. (4)sMAPE=100T∑t=1T|Ft−At|(|At|+|Ft|)/2

R2 (coefficient of determination) is a measure of how well the model explains the variance in the data; a high R2 value indicates a good fit of the model to the data (Eq. (5)). (5)R2=1−∑t=1T(At−Ft)2∑t=1T(At−y¯)2

The necessity for flexibility is forecasted by the DSO before the start of trading, which means that a longer horizon demand and generation forecasting is mandatory to estimate risks in the grid. Once gate closure time is reached and no more bids or offers can be submitted, the LFM operator clears the market and the results are sent to the DSO and aggregators (Fig. 2). However, we propose an additional step in the contracting and bidding process, right after the gate closure time and before the market clearing (Fig. 4). We suggest the DSO repeat the forecasting of demand and generation, repeat the load flow analysis, and update the flexibility required from aggregators. Forecasting the next day or forecasting the next hour can yield very different results. Thus, updating the forecast closer to real time could contribute substantially to improving forecasting performance. It is studied in detail whether the improvement of forecasting the next hour over the next day is enough to be considered as an additional step to implement in an LFM.

There is, however, a problem when it comes to forecasting the demand of individual participants of the LFMs. As per [22], the literature reveals that current forecast methodologies achieve an average forecast error of about 50% for individual household loads. This is an inherent limitation to individual household forecasting affected by human behavior.

Nevertheless, there are potential solutions to cope with this limitation. The implementation of a market mechanism with near-real-time trading of flexibility to give the possibility to compensate for forecast errors. Another solution is aggregating the participants of the LFM to reduce the combined forecast error. As mentioned, it is studied in detail how aggregating more participants of the LFM together can impact forecasting performance and help reduce the combined forecasting error. This way, however, the individual household is not the flexible asset anymore; the result of the aggregation becomes the flexible asset to manage.

As mentioned, the data used for building the forecasting models is from the FEVER project (German pilot). We consider the data points from multiple substations as an aggregation of participants in an LFM and analyze how different operations on the dataset can positively affect the performance of the forecasting algorithms. Ultimately, improving forecasting performance is going to improve the quality of the flexibility procured, as seen in [23].

The data used in the study consisted of hourly electricity consumption data from multiple substations. Data does not contain missing values or any clear biases. Before modeling, the dataset undergoes a structured feature engineering process aimed at incorporating temporal dependencies and calendar effects into the predictive variables. Other features, such as forecasted temperature or forecasted solar radiation, have not been used as inputs because they would introduce an additional source of uncertainty in the results.

The primary predictive variables are constructed by generating lagged consumption values. For each timestamp, the consumption from the previous hours was extracted at varying lags ranging from one hour up to one week (e.g., 1 to 168 h prior—a week). Temporal calendar effects were incorporated by adding the day of the week and the month as categorical predictors. Additionally, a holiday indicator is generated, and it captures exogenous demand shifts due to non-working days.

A brief analysis of data used to train the models can be seen in Fig. 5. There are multiple clear, identifiable patterns, such as weekends having load demand profiles completely different than weekdays. Conversely, weekdays seem to have very similar load demand profiles. Weekends have considerably lower load demand than weekdays. This fact is accentuated on Sundays, which is the day of the week with the lowest load demand by a substantial margin, even when compared to Saturdays.

Similarly, load demand profiles seem to be higher early in the mornings during colder months, such as January, February, March, November, and December. This fact is also seen in the last plot of Fig. 5, where the hotter quarters have a lower load demand in the mornings. This phenomenon can be attributed to an increase in the use of heating, ventilation, and air conditioning (HVAC).

In [53], a period of time between 2000 and 2016 is studied, and it is concluded that colder winters can result in occasional consumption peaks. The warmest winters over the studied period concur with a decline in energy consumption. Conversely, 2010 represented one of the coldest years, and it coincides with a hike in energy consumption compared to other years.

Thus, the use of HVAC could be one of the factors causing this fluctuation of load demand profiles during the morning part of the day.

Having too much or too few data impacts the performance of the models for STLF. Surprisingly, the more data we have does not correlate with better metric values. Table 2 shows the evaluation metrics for each set of data. The shorter training periods seem to lead to worse results, with the exception of using 1 month to train, which achieves the best sMAPE value of 4.29%. This phenomenon seems to be contrary to what is expected in machine learning, although literature explicitly states that more data does not always mean better results [38].

The best R2 value is achieved by CB using 6 months of data, with a value of 97.50%. Furthermore, the same model also achieves the lowest MAE and RMSE values, 11.24 and 14.48 kWh, respectively. Notably, adding more data does not seem to improve results starting from the 6-month dataset onward.

In most cases, CB is obtaining the best results for the evaluated metrics, with exceptions on the 2 and 3-month training periods, where LightGBM has the advantage.

Evaluating the impact of different time horizons is crucial for an LFM. Due to the technical characteristics of the grid, an LFM can be limited to using certain time horizons, such as day-ahead or hour-ahead. STLF encompasses both hour-ahead and day-ahead forecasting. Results are shown in detail in Table 3.

Unsurprisingly, the next hour time horizon achieves better results than the next day time horizon. The closer to real time approach has around 5% lower sMAPE than the day ahead approach, from 17% it goes to around 12% (depending on the forecasting method). The R2 metric also experiences a significant increase of around 12% in the hour ahead over the day ahead, from about 74% to 86%.

On another note, there is not a considerable difference between RF, CB, and LGBM algorithms. All three methods went through a process of hyperparameter tuning, and in the end, the variation between them seems to be negligible. However, LGBM takes advantage of next-day forecasting, and CB has the edge on next-hour forecasting.

The results of aggregating participants are plotted in Fig. A3 and shown in Table 4. It shows a clear improvement in all metrics as more participants are aggregated together. The average sMAPE of each participant is about 12%, while the average sMAPE of all the participants aggregated together is about 6%. The R2 metric also shows a very steady increase as the participants are aggregated.

Nevertheless, MAE and RMSE values keep increasing as we aggregate more participants; this is due to the error being the average error. In reality, the MAE error for no aggregation should be the error committed by each participant multiplied by the number of participants (3kWh×8participants=24kWh). A MAE of 24 kWh when there is no aggregation is now in line with the MAE of 13 kWh when all participants are aggregated. Table 4 summarizes all error metrics regarding the aggregation of LFM participants.

In this section, the previous results are examined in more detail, and multiple conclusions are reached from them. From the three cases we studied, there is a very clear winner when it comes to improving forecasting performance: aggregation of participants of the LFM.

Aggregating participants of an LFM seems to be really worthwhile since it increases the forecasting accuracy dramatically. Aggregating the participants together improves our sMAPE metric by about 100%, and R2 metric improves by 16% approximately. This is a sizable increase in forecasting performance just from the fact of aggregating participants together and treating them as a single unit. However, smaller aggregations such as 2-participant aggregation or 3-participant aggregation also seem to be worth considering.

Based on these results, the existence of data close to real time has a hefty impact on the accuracy of the forecasts, about 40% improvement in forecasting accuracy. This increase in forecasting accuracy could be significant enough to consider updating the flexibility request once gate closure time has been reached. This additional step, which consists of updating the flexibility request, is precisely what is proposed in Section 4.1 (Fig. 4).

On another note, training the models with half a year of data manages to obtain the best results. However, this could change in other scenarios and participants’ data. Training with half a year captures best the load consumption behaviors, but using smaller datasets, such as 1 month or 2 months, could also be considered. The accuracy of models trained with more data, such as 1 year to 5 years, does not beat the models with half a year of data, although we have seen in the literature that adding more data generally tends to positively impact the quality of the models. This could be due to the fact that load patterns change substantially over the course of the years, household appliances could have become more efficient with time, new necessities that require electricity consumption arise, households installing more and more DERs, etc.