The literature on building electricity forecasting approaches differs with regard to the envisaged timeframe. Typically, we have short-term forecasts (timeframe ranging from minutes up to 1 week), midterm forecasts (timeframe ranging from 1 week up to months), and long-term forecasts (timeframe of years), which are useful for planning infrastructure and grid investments [1], and also very long-term, policy-oriented forecasting exercises. TIMES (an acronym for The Integrated Markal Efom System) allows us to model supply and anticipated technology shifts over such a very long-term horizon, often extending as far away in time as 2100 where anticipated changes of technology are accounted for. Forecasts may also differ based on the scope of their application, which can range from a single building (to be reviewed in detail below), to a collection of buildings (e.g., neighborhood, district, etc.) [2], and up to countries [3,4,5]. The forecasting timeframe selection, the forecasting application context, and the envisaged decision support are all tightly linked together. For example, decisions related to building operational aspects are typically linked to more short-term investigations and timeframes as well as the scope of the particular building. Decisions on policy action, such as fuel taxation, may be supported by long-term timeframes and the scope of the particular country.

The overarching decision framework is that of demand response. Demand response is, in its narrow sense, about adapting consumption patterns to make the most of the pricing scheme in place [6,7]. However, the term is also used in a broader sense, whereby demand response may be any user action and response informed by electricity prices. For example, a lowering of the winter thermostat settings in time slots when prices rise is also referred to in the literature [8] as a demand response scheme; it is indeed a response, even if not a time shift response. Thus, demand response is more broadly about the user who takes action in response to changing prices.

Thus, a demand response-related forecasting timeframe must necessarily follow the respective building electricity price change timeframe. As we move to more fast-changing and flexible pricing schemes, it follows that our case is clearly that of a short-term timeframe. Indeed, in most cases, retailer prices have an hourly resolution. There are, however, also cases where the price change impact manifests over longer timeframes and might be related to cumulative consumption. For example, in the case of consumption tier pricing, prices are also determined according to the consumption levels within the billing period, which typically spans several months. In such a case, a forecast spanning a timeframe that equals the billing period also becomes pertinent.

We also need to set the timeframe in a way that is unambiguously user-friendly. Along with the above considerations, the timeframe for our investigation is set in two different and alternative ways. First, there is an hourly forecast for the next full calendar day. This means that when a user runs the forecast, she will receive 24 hourly forecasts corresponding to the consumption of the very next calendar day. Second, there is a remaining day forecast. In this alternative formulation, the user would receive forecasts for the remainder of the day. Thus, if the forecast is run at 17:15, the user would receive 6 hourly forecasts, corresponding to the six full hours remaining from 18:00 to 23:00. (Note: the forecast of 18:00 is the consumption between 18:00 and 19:00). A forecast for, literally, the next 24 h was not considered as it is essentially a subset of the two above schemes and rather less intuitive in itself. Additionally, a third timeframe that in principle needs to be considered is that of the billing period. As explained above, this may be pertinent in the case of consumption tier pricing.

Forecasting in short timeframes has been consistently addressed in the literature via a number of different approaches, both with regard to the model type as well as the parameters (features) used [9,10,11,12,13]. Approaches in the literature have evolved in two main directions. On the one hand, there are methods that rely on conventional approaches (e.g., regression, stochastic time series, ARIMA, etc.) which, due to their relative simplicity, still receive some interest in the literature. And, on the other hand, there are artificial intelligence (AI)-based methods [14,15,16,17]. Indeed, AI approaches have received increased attention and a lot of sophisticated approaches have been proposed; there is a wide consensus on the nonlinearity of the underlying phenomena in building energy, which renders AI approaches particularly pertinent.

Islam [18] has provided a comparison of these two broad approaches. In the recent decade, among AI approaches, the support vector machine (SVM) has been popular with researchers, due to the fact that it may rely on small quantities of training data. This SVM is a typical supervised learning method applied for categorization and regression, and has a solid legacy, having been introduced by Cortes and Vapnik in the 1990s [19]. It ranks high in the context of accuracy and can solve nonlinear problems using small quantities of training data. The SVM is based on the structure risk minimization (SRM) with the idea of minimizing the upper bounds of error of the object function.

More recently, hybrid methods have also been introduced by researchers [20], in an attempt to combine more than one modeling approach. Swarm intelligence (SI) approaches have been combined with ANNs as well as the SVM in search of a better forecasting accuracy. SI has been inspired by the behavior of insects and other animals, and comes in various forms such as particle swarm optimization (PSO), ant colony optimization (ACO), and artificial bee colony (ABC).

Similarly, Daut [12] has provided for an exhaustive summary of the SVM and ANN as well as hybrid approaches pursued. These authors conclude that the hybrid methods (ANN and SAI and even more SVM and SI) have shown some improvement in the performance of the accuracy of building load forecasting. Indeed, ANN has been widely utilized in different applications, but its hybridization with other methods seems to improve the accuracy of forecasting the electrical load of buildings. The same authors also expanded into reviewing the type of features typically used in the modeling exercise. Indeed, a great diversity of possible combinations have been tried out in the literature. In a categorization of 17 approaches, these authors have found that historical consumption loads are used across all of them. Diverse weather data (temperature, dry bulb temperature, dew point temperature, wet point temperature, air temperature, humidity, wind speed, wind direction, brightness of sun, precipitation, vapor pressure, global/solar radiation, sky condition) have been used in the modeling, although only a few appear consistently across the models (temperature in 11 out of the 17 cases, global/solar radiation in 9 out of the 17 cases, and humidity in 8 out of the 17 cases). Finally, indoor conditions (temperature, occupancy) as well as calendar data are also used in the modeling, albeit less frequently.

Bourdeau [21] has also provided for a review and a classification of methods for building energy consumption modeling and forecasting. He addressed both physical and data-driven approaches. As far as the latter are concerned, they also confirm the two main approaches and orientations, that is, they are mostly time series reliant as well as machine learning based. The authors analyzed 110 papers in the period between 2007 and 2019, and reported that 22 among them were based on ANN approaches, 20 on SVM approaches, 17 on time series and regression approaches, and 16 combining more than one approach. These authors referred to this last category as ensemble approaches, while reserving the term hybrid to describe the combined use of data and physical approaches. Although this represents a notable semantic difference with regard to the previous exclusively data-driven work reviewed, the results between this team and the previous one [12] converge overall. The authors also provide a thorough analysis of the features used in the modeling exercise. First in frequency is the outdoor temperature that is included in 32 of the papers reviewed; humidity and solar radiation also appear often (19 and 18 instances, respectively). Past loads are again found to be intensely used (21 instances). As to calendar data, here, the authors provide some more detailed analysis considering four types of calendar data, and, in particular, type of day (13 instances), day of the week (13 instances), time of the day (12 instances), and month of the year (8 instances). Occupancy data appear in 7 instances while indoor temperature only in 1 instance (although other types of indoor data appear in 2 more instances). Overall, there seems to be a good convergence with the related analysis of feature frequency presented by the previous researchers. This is also confirmed by Nti [22], who examined the various aspects of electricity load forecasting. The findings indicated that weather factors are employed in 50% of the electricity demand predictions, while 38.33% rely on historical energy usage. Additionally, 8.33% of the forecasts considered household lifestyle, with 3.33% taking even stock indices into account.

All approaches reviewed above aim at reducing the forecasting error, by means of the typical error metrics used: the Root Mean Squared Error (RMSE), the Mean Absolute Error (MAE), and the Mean Average Percentage Error (MAPE). Thus, accuracy is typically the sole optimization criterion.

In time, however, it became apparent that there are also other important issues besides accuracy related to the forecasting. In particular, explainability considerations, implying the ability to understand why the forecast works the way it does and get this understanding over to the user, have started receiving increased attention over the last few years. The need to study and analyze the possible tradeoff between accuracy and explainability emerges now in a natural way and is something we will also engage with in the following.

In the energy domain, Mouakher [23] introduced a framework seeking to provide explainable daily, weekly, and monthly predictions for the electricity consumption of building users of a LSTM-based neural network model, based on consumption data as well as external information, and, in particular, weather conditions and dwelling type. To this extent, in addition to their predictive model, the authors also developed an explainable layer that could explain to users the particular, every time, forecast of energy consumption. The proposed explainable layer was based on partial dependence plots (PDPs) which unveiled what happens to energy consumption whenever a feature changes while keeping all other features constant.

Kim [24] attempted to explain the feature dependence of a deep learning model by taking account of the long-term and short-term properties of the time series forecasting. The model consisted of two encoders that represented the power information for prediction and explanation, a decoder to predict the power demand from the concatenated outputs of encoders, and an explainer to identify the most significant attributes for predicting the energy consumption.

Shajalal [25] also embarked on the task of energy demand forecasting and its explainability. The focus was on household energy consumption, for which they trained an LSTM-based model. For the interpretability component, they combined DeepLIFT and SHAP, two techniques that explain the output of a machine learning model. This approach enabled the illustration of time association with the contributions of individual features.