Enhancing Short-Term Electrical Load Forecasting for Sustainable Energy Management in Low-Carbon Buildings

Abstract

Accurate short-term forecasting of electrical energy loads is essential for optimizing energy management in low-carbon buildings. This research presents an innovative two-stage model designed to address the unique challenges of Electricity Load Forecasting (ELF). In the first phase, robust data preprocessing techniques are employed to handle issues such as outliers, missing values, and data normalization, which are common in electricity consumption datasets in the context of low-carbon buildings. This data preprocessing enhances data quality and reliability, laying the foundation for accurate modeling. Subsequently, an advanced data-driven modeling approach is introduced. The model combines a novel residual Convolutional Neural Network (CNN) with a layered Echo State Network (ESN) to capture both spatial and temporal dependencies in the data. This innovative modeling approach improves forecasting accuracy and is tailored to the specific complexities of electrical power systems within low-carbon buildings. The model performance is rigorously evaluated using datasets from low-carbon buildings, including the Individual-Household-Electric-Power-Consumption (IHEPC) dataset from residential houses in Sceaux, Paris, and the Pennsylvania–New Jersey–Maryland (PJM) dataset. Beyond traditional benchmarks, our model undergoes comprehensive testing on data originating from ten diverse regions within the PJM dataset. The results demonstrate a significant reduction in forecasting error compared to existing state-of-the-art models. This research’s primary achievement lies in its ability to offer an efficient and adaptable solution tailored to real-world electrical power systems in low-carbon buildings, thus significantly contributing to the broader framework of modeling, simulation, and analysis within the field.

1. Introduction

The notion of energy prediction for resources planning and management has been investigated since 1966, and it is not a new topic [1]. The first stage in organizing, allocating, investing in, and managing national or local grids is always to develop a precise and trustworthy technique of Electricity load forecasting in order to ease the functioning, transfer, retention, and production of energy [2,3,4,5,6]. ELF error minimization can have significant cost- and power-saving advantages [7]. For instance, a study of 19 organizations found that reducing prediction error by 1% can result in energy savings of up to 10,000 megawatts, which means an efficient ELF system could conserve 1.6 million USD annually [8]. This figure might exceed millions of dollars if we take into account data of electricity consumption on a country level. Since the demand for power energy is quickly rising, current research mostly concentrates on constructing an effective model to enhance forecasting effectiveness [6].

According to a report by the US Energy Information Administration, the rise in electricity power consumption is likely to increase up to 28% from 2015 to 2040 [9]. According to a different estimate by the International Energy Agency, about 36% of electricity power generated is used by Demand Side Management (DSM), which is economically significant in the context of something like low-carbon economics [10]. Therefore, in such a situation, a precise Electricity Load Forecasting model is essential for energy preservation in situations regarding building construction and refurbishment. Additionally, this can make it easier to control energy surges in big, multipurpose buildings [11]. Appropriate ELF may be created through key spots and aids, which would assess both the residential and commercial electricity distribution sectors. For instance, it is crucial to point out the exact indicators that have a significant impact over the consumer state, prior to making a prediction, as well as to incorporate such variables in the prediction models. Meter-reading sensors produce results regarding power use, and these data are required to be processed, since they contain outliers, incomplete numbers, and unusual consumer behavior. Additionally, the nonlinear interactions between the input and output parameters of electricity power consumption make ELF extremely difficult; as a result, it is crucial to use a suitable modeling approach.

According to information from a recent survey, ELF is categorized into the following three categories: short-term, long-term, and medium-term ELF [12]. In long-term electricity load forecasting, the predictions are carried out for a year or more than one year ahead, which is very important for networks as relates to the future planning and forecasting of electricity [13,14]. In medium-term electricity load forecasting, predictions are conducted for a period of one month up to a year, and these are used for the purposes of maintaining availability and stability, which electricity power generation needs [3,15]. Short-term ELF is a type of forecasting in which predictions are made from a duration of mere minutes to a week ahead, and these can aid in organizing economical deployment and the best engagement of electricity generation and load; these are then used for adjusting and controlling price, consumption demand, and load in real time [15]. Yazici et al. [16] analyzed real-world electricity load data using one-dimensional CNNs, GRU [17], and LSTM variants. These models are used to address critical short-term load forecasting tasks, specifically focusing on predicting load values for 1 h and 24 h into the future. Bashir et al. [18] proposed a hybrid model, named prophet-LSTM, for short-term electricity load forecasting. Similarly, a heuristic configuration-based deep learning model is used for short-term electricity load forecasting, which demonstrates forecasts for the load of two different commercial buildings at periods of 24 h and 30 min into the future [19]. F. Alsharekh et al. [20] proposed R-CNN and ML-LSTM, with some data preprocessing steps, for electricity load forecasting. They used commercial PJM and residential IHEPC datasets and reduced error rates compared to existing methods. Furthermore, transformer-based models with a modification of the NLP transformer workflow, the new handling of contextual features, and the addition of N-space transformation for load forecasting is presented in [21]. Chai et al. proposed a new method for ELF, which is based on GRU-TCN, VWM, and hybrid models and achieves optimal performance. Despite the fact that each ELF prediction has their own application, short-term electricity load forecasting drew a great amount of attention in the research community due to its ample number of functions, such as the real-time control and adjustment of the electricity power load, revenue maximization, and management [22]. Due to the above-mentioned reasons, our focus in this work is only on short-term electricity load forecasting predictions.

Electricity load forecasting is crucial for establishing efficient power provision services, as precise prediction is difficult because of noisy data, which leads to inaccurate forecasts [23]. To develop an electricity load prediction model, various models like conventional machine learning in addition to deep learning approaches have been used for electricity load forecasting; nevertheless, these approaches are unreliable and produce poorly predicted values [24]. Some significant obstacles faced by these approaches include training and learning from scratch, overfitting, short-term memory, and failure in the complex correlation of variables. Additionally, some researchers have used a combination of models in response to the shortcomings of the use of a single model, because the individual use of models results in failure to learn the spatiotemporal features of data at the same time; as a result, this field offers opportunities for new research and analysis.

In this research, the main objectives are twofold: first, to develop a robust and accurate short-term electricity load forecasting (ELF) model; and second, to demonstrate the practical applicability of a data-driven approach in addressing challenges related to outliers, missing values, and the complex spatiotemporal features inherent in electricity consumption data. The primary aim is to contribute to the field of ELF by introducing a two-stage model that combines data-cleansing techniques with an innovative architecture, featuring a residual Convolutional Neural Network (CNN) and an Echo State Network (ESN). This approach seeks to overcome the limitations of traditional forecasting models, such as unreliable predictions and difficulty in capturing complex correlations among variables. The need for this research arises from the critical role of accurate ELF in efficient power provision services, with potential cost- and energy-saving advantages. Its novel contribution lies in the development of a data-driven model tailored for short-term ELF, addressing the challenges posed by noisy data and complex variable interactions. The research’s significance extends to its practical application in real-world scenarios, as evidenced by the evaluation of diverse datasets representing residential and regional electricity consumption. Through these objectives and contributions, this study aims to advance the scientific understanding and practical implementation of data-driven decision-making in the realm of electrical power systems. The main achievements of our research are concisely explained as follows:

• Robust Data Preprocessing for Enhanced Reliability: Robust data preprocessing techniques are implemented to mitigate ambiguity arising from outliers, missing values, and data normalization issues commonly found in electricity consumption datasets within the context of low-carbon buildings. These techniques enhance data quality and reliability, providing a solid foundation for accurate load forecasting.
• Advanced Data-Driven Modeling for Low-Carbon Buildings: This research introduces an innovative hybrid architecture that combines a novel residual Convolutional Neural Network (CNN) with a layered Echo State Network (ESN) to capture both spatial and temporal dependencies in the data. This advanced modeling approach is specifically designed to improve load forecasting accuracy while addressing the unique challenges posed by electrical power systems in low-carbon buildings.
• Real-World Testing and Superior Performance in Low-Carbon Buildings: Comprehensive testing across multiple datasets, including the IHEPC and PJM datasets, as well as evaluations in ten distinct regions within the PJM dataset, demonstrates the model’s robustness and its capacity to outperform existing state-of-the-art models in the context of low-carbon buildings. This achievement underscores the efficiency and adaptability of our system in real-world electrical power system forecasting scenarios within low-carbon building environments.

The rest of the paper is structured as follows: In Section 2, we survey prior work on ELF and present an overview of related research. The methodology proposed is detailed in Section 3. The results of our analysis are then discussed in Section 4. Finally, in Section 6, we conclude the manuscript.

2. Related Work

From the last few decades, researchers have concentrated on various methods to improve and enhance the performance of electricity load forecasting. On a learning basis, there are four primary mainstream approach categories: the first one is physical, the second one involves a persistence learning technique, the third one is statistical, and the last one involves an artificial intelligence approach [25]. Physical models are dependent on mathematical formulas that use historic and environmental data. Because of their significant memory consumption and high computing time requirements, physical models are not perceived as reliable for the prediction of electricity load forecasting [26]. The persistence learning technique is the simplest approach for the future prediction of sequential data observance, like electricity load forecasting, but it fails in long-time data prediction [26], making it inappropriate for ELF. Statistical models are less expensive in terms of computational complexity compared to physical models [27], and these models are mostly dependent on autoregressive techniques, like linear regression, Generalized-Autoregressive Conditional-Heteroskedasticity (GARCH) [28], and Autoregressive-Integrated Moving-Average (ARIMA) [29] methods. As statistical approaches are not capable of directly extracting uncertainty trends from electric data, they can therefore be used in combination with different models to mitigate uncertainty [30]. On the contrary, approaches that use AI are built on shallow and deep-ordered procedures and can therefore learn complicated and nonlinear structures. Nonetheless, some shallow-based approaches have not worked well for feature mining like Artificial Neural Networks (ANNs), Support Vector Regression [31,32], Random Forest (RF) [33], Wavelet Neural Networks [34], and the Extreme learning Machine model. Therefore, to increase the performance of these shallow-based models, they require many feature selection and feature extraction methods, which present a challenging task.

AI-based approaches are further divided into two types: the first includes traditional machine learning models, and the second involves deep learning. On the contrary, demand side management can address the consideration of shallow-based approaches through the learning of multi-layer structural features from the input data given. CNNs [35] and Long Short-Term Memory (LSTM) [36,37] are presently the most effective architectures for the study of temporal data, whereas electricity load forecasting (ELF) has both types of features, e.g., spatial and sequential features. Furthermore, CNNs are unable to learn the sequential features of electricity time series data [38]. Therefore, in order to address this issue, many hybrid architectures have been devised to precisely predict electricity load forecasting [39]. In the work of [40], a CNN was used in combination with LSTM for short-term electricity load forecasting. In [41], Seasonal ARIMA (SARIMA) was integrated with an LSTM model for electricity consumption prediction. In another work by [23], an empirical setting with a 10-fold evaluation was used to integrate both a CNN and a bidirectional long short-term memory model in order to increase accuracy. Khan et al. [42] created a convolution neural network with an LSTM autoencoder prediction of short-term electricity load forecasting. Likewise, in [25,43], a hybrid model was proposed in which two models, a CNN and Gated-Recurrent-Unit (GRU), were integrated. Furthermore, other popular electricity load forecasting models have been created, such as the multivariable Experimental Mode Decomposition network, an expanded CNN in combination with LSTM [44], hybrid artificial neural network models [20,45], Empirical Mode Decomposition (EMD), and Extreme Learning Machine (ELM) models [46]. These approaches overcome the issues of shallow-based models, but they are unreliable for accurate implementation due to low forecasting accuracy.

In summary, each type of model has some flaws. For example, persistence approaches perform well for short-time predictions but perform badly for long-sequence predictions. Such techniques are useful for predicting atmospheric data, but they need long sequences for best results, which then become computationally complex and need more resources [47]. Moreover, these types of models underperform when abrupt changes appear in the input data sequence; therefore, researchers do not suggest such models for electricity load forecasting [48]. Statistical methods work efficiently with linear data, but they fail in the learning of nonlinear, complicated trends in electricity data, demonstrating their ineffectiveness for large sequence prediction in electricity consumption data [49]. Consequently, these approaches are also not suggested for electricity load forecasting in the most current research. Additionally, AI-based techniques need manually extracted features, and they can only learn one type of feature at a time: either the spatial or temporal. However, ELF problems have both types of features, spatial as well as temporal. Because of these shortcomings, such techniques are also not suggested for electricity load prediction [25]. Hybrid approaches are able to extract spatiotemporal characteristics and acquire nonlinear and complicated patterns, which allow them to solve these shortcomings. However, predicting accuracy remains a difficult problem that necessitates additional attention from researchers.

Hence, in this research, we provide a two-phase framework for short-term electricity load forecasting. The first step involves removing outlier records from electricity consumption data, filling missing values, normalizing data, and down sampling input data. Then, preprocessed data are given to a deep residual Convolutional Neural Network model for the extraction of important features; after this, the data are raised forward to ESN layers to obtain the sequential features of electricity consumption data. Then, the proposed model uses a residual CNN with an Echo State Network (ESN) for the first time [50] for electricity consumption prediction (ECP). The residual block then supports our model to become stable for short-term sequences and does not become stuck in local minima nor overfitted. In short, our proposed residual CNN model with a layered ESN paradigm addresses issues with shallow-based architecture and deep learning models, demonstrating enhanced accuracy compared to existing state-of-the-art (SOTA) hybrid architectures.

3. The Proposed Model

In order to provide end customers with an adequate energy supply, electricity consumption prediction is very important. Due to changing customer usage, noisy data layout, and uncertain weather patterns, accurate electricity consumption prediction is a difficult task. As described in Section 2, different methods are proposed for this reason to forecast power consumption. However, for a reliable Micro grid system, prediction accuracy still requires much improvement. Therefore, this study proposes a new framework combining a CNN and ESN for electricity consumption prediction (ECP), and a high-level framework is given in Figure 1. The architecture of the CNNESN model consists of two sub-networks: the first one is the Convolution Neural Network (CNN), and the second one is the Echo State Network (ESN); these two sub-networks are described in the next subsections.

3.1. Convolutional Neural Network

Scholars in the fields of computer vision [51,52], image recognition [53,54], and video data analysis [55,56,57,58] are presently interested in the results of new deep learning models, particularly the CNN, a branch of AI that takes its primary inspirations from the human vision system [59]. Because of sharing weights and the internal connection approach, the CNN model has demonstrated impressive results in a variety of applications, including energy forecasting, load management prediction, and numerous others. In general, CNN architecture has three types of layers: the first one is the convolutional layer, the second one is the pooling layer, and lastly, a fully connected layer is used in the network [60]. The convolution layer in the CNN architecture retrieves the particular features from the input data, which is then convolved through a filter, and the resulting extracted feature from the previous layers of the network is given the final feature vector i. The initial layer of the network extracts the local feature, which is then combined to form global features with the use of the intermediate layers of the network. Mathematically, we can represent the convolution layer of the network as follows:

$${ O}_{ i}^{\left(l\right)}=\left({\sum }_{jϵc_i}^{t_j^{l-1}}\ast W_{j.i}^{\left(l\right)}\right)+b_i^{\left(l\right)}$$

(1)

$${ F}_i^l=f\left(y_i^l\right)$$

(2)

where Fli is the feature vector extracted through the lth convolution layer. The convolution’s outcome is indicated by O(l)I; its bias factor is indicated by b(l)i, while the convolution filter is indicated by the letters W(l)j.i, and the activation function is indicated by the f letters. The activation function used in this work is the Rectified Linear Unit (ReLU); its mathematical equation is shown below in Equation (3).

$$f\left(x\right)=max(0,x)$$

(3)

To lower the contiguous quality of the given input historical data, pooling layers are mostly used, and various kind of pooling layers exist, such as max pooling layers, min pooling layers, and average pooling layers. Here in our work, we have chosen max pooling for selecting eminent values from data. Recent advancements in the field of image recognition have shown the efficacy of CNN models. However, as the depth of the network increases, the model’s accuracy becomes saturated, leading to a problem known as degradation. To mitigate this issue, residual learning architecture, known as ResNet, was introduced. Unlike traditional CNNs, which directly learn the target function, residual learning defines the target function as H(a) = F(a) + a, allowing for the creation of deeper networks. By using this residual strategy, more complex relationships between the input and output are captured. The features extracted from the residual CNN can then be considered encoded features, representing historical data for hourly electricity load forecasting. In this work, a deep residual CNN is utilized to effectively extract spatial features from electricity data. The output is then passed onto an ESN, which models the temporal information of the historical electricity data.

3.2. Echo State Network

Deep learning models with different layers of architecture have recently attracted much interest in the area of artificial neural network models [61]. The stratified standard Recurrent Neural Network has also proven to have an important role in a variety of challenging operations, like deep learning. The work of Correll et al. [62] initially combined the Echo State Network (ESN) with deep learning architecture, which is computationally inexpensive in comparison to previous RNN types, because ESN is a new and unique type of Recurrent Neural Network [63]. ESN was designed by Jaeger et al. [64] and offers an important framework, as well as a supervised deep learning network technique, to the Recurrent Neural Network; in the Echo State Network, the hidden layers are formed from reservoirs, whereas the basic components of the ESN design include a reservoir designated by the letters ‘R’, an input designated by the letters ‘I’, and an output designated by the letters ‘O’. The incoming input as well as the internal and resultant output units, together with the accompanying computations, are listed in Equations (4)–(6). Following that, the internal processing and output processing unit update equations are listed in Equations (7) and (8).

$$u\left(i\right)={\left[u_1\left(i\right),u_2\left(i\right)\cdots .u_l\left(i\right)\right]}^k$$

(4)

$$x\left(i\right)={\left[x_1\left(i\right),x\left(i\right)\cdots .x_R\left(i\right)\right]}^k$$

(5)

$$y\left(i\right)={\left[y_1\left(i\right),y\left(i\right)\cdots .y_o\left(i\right)\right]}^{k }$$

(6)

$$x\left(i+1\right)=f\left(W_{in}\ast u\left(i+1\right)+W\ast x\left(i\right)+W_{back}\ast y\left(i\right)\right)$$

(7)

$$y\left(i+1\right)=g\left(W_{out}\ast u(i+1)\right)$$

(8)

$$W_{out}={\left(M^{-1}\ast T\right)}^k$$

(9)

where the ‘f’ and ‘g’ in Equations (7) and (8) shows the activation function of both the reservoir and the output units, while the total feature metrics comprise ‘WInput(R*I)’, ‘Wback(R*O)’, ‘WRes(R*R)’, ‘WOut(O*R)’, input value, backward output, reservoir, and readout metrics correspondingly. The reservoir’s weight vector changes during the process of learning, whereas ‘Wout’ is chosen at random and remains fixed [65]. For calculating the readout weights, the reservoir state metrics ‘T ((S – So + 1) ∗ O)’ and the target outputs ‘I ((S – So + 1) ∗ R)’ are chosen. Here, ‘S’ shows the training stride, whereas ‘So’ shows the washout or erasing time stride. When the time step is greater than or equal to ‘Vo’, the readout values are calculated by Equation (9). These reservoirs serve as the primary function in ESN, since they have a significant impact on the network’s overall quality, with the help of three key characteristics. Additionally, the quantity of the reservoir neuron ‘N’ has demonstrated a significant influence on the functionality of ESN because of its inner structure, which is connected to knowledge in the internal hidden state [66]. Furthermore, it relies on the volume of input training examples and the difficulty of the desired outcomes. Accordingly, the absolute eigenvector of the weight vectors ‘W’; the particular radius ‘q’, when ‘q’ is specified within the 0 and 1 periods; and the rate of connection ‘a’ all have an impact on ESN functionality. Thus, the conclusion is that the ESN is faster and better than simple RNNs. In conclusion, the ESN performs learning and prediction tasks quicker than and superior to RNNs [66].

3.3. CNNESN Architecture

From the historical data of electricity power consumption, the outlier value, duplicate value, and the missing data were removed through preprocessing. Outlier values in data mostly occur due to unconditional weather situations, electricity short circuits, sensor faults, etc. [67]. For the removal of such outlier values from data, the three-sigma rule was used in this work [68]; here, the mathematical formula used to obtain the three-sigma rule is shown in Equation (10):

$$f\left(d_i\right)=\left\{\begin{array}{c}avg\left(D\right)+2std\left(D\right),if{d}_i>avg\left(D\right)+std\left(D\right),\\ di,otherwise\end{array}\right.$$

(10)

Here, D represents data, the average is represented by avg(D), and the std(D) represents the standard deviation of the data. This data preprocessing approach considers the impact of outliers. We employed the three-sigma rule to detect and address outliers, replacing them with values within two standard deviations from the mean. This strategy was chosen to balance the retention of potentially informative outlier data points while preventing them from unduly affecting model predictions. We believe this approach helped to maintain data integrity while significantly reducing the risk of outliers influencing our forecasting model. For missing value recovery, the NAN interpolation method was used, whose mathematics are shown in Equation (11).

$$f\left(d_i\right)=\left\{\begin{array}{c}\frac{d_{i-1}+d_{i+1}}{2},d_i\in NAN,d_{i-1},d_{i+1}\notin NAN\\ 0,a_i\in NAN,a_{i-1} {or a}_{i+1}\in NAN ,\\ a_i,a_i\notin NAN\end{array} \right.$$

(11)

where di represents electricity power consumption data. If any of these values were empty, we changed it with NAN. Additionally, we transformed enormous, diverse samples into a certain range using the normalization approach. Because of the diverse range of electricity power consumption, the normalizing approach was used. We employed the min–max normalization approach with a range of (0–1).

In our data preprocessing phase, we applied robust techniques to ensure the quality and suitability of our dataset for electricity load forecasting in low-carbon buildings. Outliers, which can arise from various factors such as extreme weather conditions, electricity short circuits, or sensor faults, were addressed using the three-sigma rule. This rule detects outlier values and replaces them with values within two standard deviations from the mean. The rationale behind this choice was to retain potentially informative outlier data points while mitigating their undue influence on model predictions. To handle missing values, we employed the NAN interpolation method, which estimates missing data based on neighboring values, ensuring data completeness. We also utilized the min–max normalization approach to bring diverse samples into a common range within the (0–1) range. These preprocessing techniques were carefully selected to enhance data quality and prepare it for robust model training.

For training, validation, and the testing of the CNNESN system, the refined and normalized data were fed and used for the prediction of electricity power consumption. Historical data were used, which were preprocessed and separated into 70% training data, 10% for validation purposes, and 20% for testing purposes. During training, the refined data were given as input to the proposed residual CNN network layer, which extracted spatial features and patterns. After this, the ESN network model learned the temporal or sequential information from the given data. The resultant output vales of the ESN are then fed to the dense layers for the final ECP. In the proposed CNNESN model, there were four convolutional layers with the filter sizes of 16, 32, 64, and 128 with the kernel sizes of 7, 5, 3, and 1 with the ReLU activation function, whereas the Echo State Network (ESN) had one reservoir with 64 units where the activation function of tanh was used. Finally, two dense layers of the sizes 128 and 60 were applied for final forecasting. A general working flow from input to output prediction is given in Figure 2.

4. Results

The proposed residual CNN with ESN architecture was evaluated using IHEPC [69] and PJM [70] datasets. The experiments were conducted on an Intel Core i5 CPU with 12 GB RAM and an 8 GB GeForce RTX 1060 GPU using Windows 10. The training was performed using Python 3.7, Keras framework with TensorFlow backend, with an initial learning rate of 0.001, Adam optimizer, and batch size of 256. The performance was evaluated for both hourly and daily ELF.

4.1. Evaluation Metrics

The performance of the proposed network was evaluated using the MAE, MSE, RMSE, MAPE, and NRMSE metrics. These metrics assessed the accuracy of the model’s predictions and the closeness of the actual values to the model’s outputs. The MAE measured the difference between actual and predicted values, while MSE calculated the square of the prediction error. RMSE represented the residual between the actual values and the model’s outputs. The mathematical equations for each metric is given in [10].

4.2. Dataset Description

The proposed model was evaluated on the IHEPC and PJM benchmark datasets. The IHEPC dataset contains nine variables including date, time, global active and reactive power, voltage, global intensity, and three submetering variables. The details of these variables and their units are briefly described in [10]. The dataset is based on actual electricity consumption readings in a residential building in France, collected over a four-year period from 2006 to 2010. It consists of per-minute electricity consumption data with a total of 2,075,269 instances and includes over 35,000 missing and outlier values.

The PJM Interconnected organization operates the Eastern Interconnection electricity grid in the United States and transmits electricity to 14 US districts, including Maryland, Delaware, and Michigan. The power consumption data used in the experiments were obtained from the official PJM website and recorded with a one-hour resolution in MW for selected dates. Ten regions were selected from the PJM dataset for our experiments for a fair evaluation of the proposed model with baselines, whereas the statistical information for both the IHEPC and PJM datasets are given in [10].

4.3. Assessment Using the IHEPC Dataset

The superiority of our model on the IHEPC dataset was demonstrated through evaluations of hourly and daily ELF data. The proposed model achieved high accuracy in performance tests. The results for the hourly ELF evaluation are presented in Figure 3a, and the daily load forecasting results are displayed in Figure 3b.

Table 1. Performance comparison of R-CNN and ML-LSTM with the state-of-the-art, hourly IHEPC dataset.

Method	MSE	MAE	RMSE	MAPE	NRMSE
Linear regression [71]	0.4046	0.4176	0.6361	74.52	-
CNNLSTM [71]	0.3738	0.3493	0.6114	31.84	-
SE-AE [72]	0.384	0.395	-	-	-
CNN Stacked LSTM [10]	0.038	0.003	0.058	-	7.579
FCRBM [73]	-	-	0.666	-	-
CNN-GRU [24]	0.22	0.33	0.47	-	-
Residual GRU [74]	0.1753	0.2635	0.4186	-	-
Multilayered LSTM [20]	0.0011	0.0144	0.0325	1.024	-
CNNLSTM-autoencoder [23]	0.19	0.31	0.47	-	-
ANN [76]	-	1.08	1.15	-	-
GRU [77]	0.17	0.19	0.41	0.6	-
CNN-BiGRU [78]	0.18	0.29	0.42		0.2341
ESN [15]	0.0022	0.0266	0.0472	-	-
STLF-Net [79]	0.1924	0.2674	0.4386	36.24	-
CNN [10]	0.37	0.47	0.67	0.76	-
Proposed	0.0007	0.0088	0.0257	0.2945	0.0024

compares the performance of the proposed model with other baseline methods for hourly ELF.

Table 2. Performance comparison of the proposed model with the state-of-the-art, daily IHEPC dataset.

Method	MSE	MAE	RMSE	MAPE	NRMSE
FCRBM [73]	-	-	0.828	-	-
CNN-LSTM [71]	0.103	0.256	0.322	32.83	-
LSTM [71]	0.240	0.412	0.490	38.72	-
Linear regression [71]	-	0.502	-	52.69	-
CNN [42]	0.006	0.05	0.07	0.69	-
Multilayered LSTM [20]	0.002	0.0132	0.0447	2.457	-
Proposed	0.0011	0.0194	0.0379	0.3421	0.0027

provides a comparison of daily ELF results. For hourly load forecasting, the proposed model is compared with Linear regression [71], CNNLSTM [71], SE-AE [72], CNN Stacked LSTM [10], FCRBM [73], CNN-GRU [24], Residual GRU [74], Multilayered LSTM [20], CNNLSTM-autoencoder [23], CNN-BDLSTM [75], ANN [76], GRU [77], CNN-BiGRU [78], ESN [15], STLF-Net [79], and CNN [42]. By comparison, the worst performance was achieved using Linear regression [71], and a better performance was achieved using Multilayered LSTM [20]. However, the proposed model further reduced the error scores and achieved the best results with the lowest error rates: 0.0088 for MAE, 0.0007 for MSE, 0.0257 for RMSE, 0.2945 for MAPE, and 0.0024 for NRMSE.

The proposed model was evaluated on daily ELF data, and it was compared to several baseline models. The results showed that the proposed model achieved the lowest error rates in terms of MAE, MSE, and RMSE. Specifically, the MAE, MSE, RMSE, MAPE, and NRMSE error rates were 0.030, 0.015, 0.120, 0.3421, 0.0027, respectively. These results indicate that the proposed model provides highly accurate predictions compared to the baseline models, including FCRBM [73], CNN-LSTM [71], LSTM [71], Linear regression [71], CNN [42], and Multilayered LSTM [20]. These results demonstrate the superiority of the proposed model in terms of daily ELF data prediction, making it a valuable tool for forecasting electricity consumption.

4.4. Assessment Using the PJM Dataset:

The performance of R-CNN with ML-LSTM was evaluated on several datasets for daily load forecasting in the PJM region. Out of the 14 datasets available, we chose 10 as selected by a previous study. Our proposed model was compared with existing models and was found to have the lowest error rate for each dataset, as shown in

Table 3. Comparative analysis of PJM dataset.

Dataset	Method	RMSE	MAE	MSE	MAPE	NRMSE
AEP	Mujeeb et al. [81]	0.386	-	-	4.32	0.08
	Gao et al. [80]	0.49	-	-	7.98	-
	Han et al. [77]	0.054	-	-	-	-
	Khan et al. [10]	0.031	0.001	0.027	-	7.169
	Alsharekh et al. [20]	0.0223	0.0163	0.0005	0.5504	-
	Proposed	0.0010992	0.0001174	0.0000012	0.3475	0.0004
DAYTON	Khan et al. [10]	0.046	0.033	0.002	-	7.955
	Alsharekh et al. [20]	0.0206	0.0144	0.0004	0.4982	-
	Proposed	0.001541	0.0002197	0.0000024	0.3144	0.0014
COMED	Khan et al. [10]	0.044	0.030	0.002	-	6.892
	Alsharekh et al. [20]	0.0216	0.0131	0.0005	0.5475	-
	Proposed	0.0015333	0.0000876	0.0000024	0.2987	0.0016
DOM	Khan et al. [10]	0.057	0.039	0.003	-	8.131
	Alsharekh et al. [20]	0.0212	0.0138	0.0005	0.5987	-
	Proposed	0.0010562	0.000062	0.0000011	0.2647	0.0009
DEOK	Khan et al. [10]	0.053	0.036	0.003	-	6.830
	Alsharekh et al. [20]	0.0174	0.0129	0.0003	0.3974	-
	Proposed	0.0013028	0.0000678	0.0000017	0.3244	0.0021
EKPC	Khan et al. [10]	0.055	0.034	0.003	-	6.667
	Alsharekh et al. [20]	0.0274	0.0202	0.0008	0.7965	-
	Proposed	0.0012136	0.0001617	0.0000015	0.3387	0.0027
DUQ	Khan et al. [10]	0.054	0.041	0.003	-	6.724
	Alsharekh et al. [20]	0.0430	0.0277	0.0009	0.8234	-
	Proposed	0.0012856	0.0001373	0.0000017	0.2588	0.0031
PJME	Khan et al. [10]	0.043	0.031	0.002	-	5.820
	Alsharekh et al. [20]	0.0199	0.0128	0.0004	0.4721	-
	Proposed	0.0017765	0.0002533	0.0000032	0.3178	0.0018
NI	Khan et al. [10]	0.050	0.033	0.002	-	6.228
	Alsharekh et al. [20]	0.0178	0.0129	0.0003	0.3748	-
	Proposed	0.0017647	0.0001008	0.0000031	0.3422	0.0011
PJMW	Khan et al. [10]	0.038	0.027	0.001	-	7.381
	Alsharekh et al. [20]	0.0145	0.0102	0.0002	2864	-
	Proposed	0.0012382	0.0000727	0.0000015	0.3255	0.0015

. The comparison included studies by Gao et al. [80], Mujeeb et al. [81], Khan et al. [10], Chou et al. [82], Han et al. [77], and Alsharekh et al. [20]. The superior performance of our proposed model was demonstrated in all datasets. Figure 4a displays the prediction results of the AEP region, while Figure 4 demonstrates the prediction results of other PJM regions.

4.5. Validating Robustness on the AEP Dataset

In addition to the comprehensive evaluation using the IHEPC and PJM datasets, we sought to validate the robustness of our proposed model using the Appliances Energy Prediction (AEP) dataset [83], which provides an additional layer of diversity in terms of data characteristics. This dataset, recorded with a high temporal resolution of ten seconds in a house near Mons city, Belgium, over a span of 4.5 months, presents unique challenges for load forecasting.

Table 4. Comparative analysis of the proposed model with the state-of-the-art AEP dataset.

Method	RMSE	MAE	MSE
LSTM [84]	0.4562	0.3103	0.2081
Transformer [85]	0.3021	0.2819	0.0913
GRU [86]	0.3913	0.3034	0.1531
CNN-LSTM [71]	0.4390	0.3142	0.1927
Proposed	0.2670	0.1978	0.0713

provides a comparative analysis of our proposed model with state-of-the-art methods on the AEP dataset. Notably, our model exhibits exceptional performance with significantly lower values for all of the evaluation metrics, including RMSE, MSE, and MAE, in comparison to existing models.

These results underscore the robustness and adaptability of our proposed model across diverse datasets. It excels not only in the context of low-carbon buildings but also in high-resolution, short-term forecasting tasks in a residential setting, as demonstrated by its superior performance on the AEP dataset. This validation highlights the model’s versatility and its potential to significantly contribute to the field of electricity load forecasting across various data scenarios.

5. Discussion

The integration of a residual CNN and an ESN in our approach signifies a shift from conventional methodologies, reflecting a profound understanding of the complexities involved in short-term electrical demand prediction. The approach proposed captures spatial and temporal interdependencies, comprehensively resolving the dynamic features inherent in the network. Leveraging the spatial feature extraction capabilities of CNNs and the temporal pattern-capturing ability of ESNs, our model offers a complete framework that accurately forecasts short-term electrical demand. The actual litmus test for any forecasting model lies in its adaptability to diverse scenarios.

The importance of adequate data preprocessing methods serves as a solid foundation for ensuring dependable load forecasting. The precise inclusion of outliers, treatment of missing values, and normalization procedures are critical steps ensuring feature refinement before ultimate predictions. This thorough data purification process enhances precision and instills confidence in the outcomes, a crucial element in decision-making for low-carbon buildings. Rigorous assessments on the IHEPC and PJM datasets, spanning various locations within the PJM dataset, validate the robustness and practical applicability of our proposed model. In addition to the IHEPC and PJM datasets, we tested the adaptability of the proposed model using the AEP dataset, which adds data variety. This dataset, collected over 4.5 months in residence in Mons, Belgium, with a ten-second temporal resolution, provides unique load forecasting challenges. Compared to existing models, our model performs well with reduced RMSE, MSE, and MAE values. Our approach is stable and adaptable across varied datasets, as shown by these findings. Its persistent outperformance against existing methodologies underscores its effectiveness in navigating the complexities of low-carbon building settings. This adaptability positions our model as a significant asset for decision makers in forecasting complex electrical power systems.

Beyond technical complexities, our study highlights concrete energy-saving ramifications. The cost- and energy-saving benefits illuminated in the introduction find tangible expression in the capacity of our proposed model to provide precise load predictions. This technique plays a pivotal role in broader endeavors to achieve sustainable and economically feasible energy consumption by facilitating more efficient energy use through accurate forecasts. Acknowledging the obstacles faced by conventional machine learning and deep learning methodologies enhances the comprehensiveness of our discourse. From challenges in training to concerns related to overfitting and short-term memory, these obstacles are not merely recognized but actively addressed, albeit gradually. The introduction of our two-stage data-driven model serves as a solution to these challenges, demonstrating a synergistic effect that overcomes the constraints of individual models. The integration proposed enriches the discourse and positions our study as a crucial addition to the progression of forecasting approaches within the complex domain of short-term electrical demand prediction.

6. Conclusions

In conclusion, our research introduces a robust framework for short-term ELF that traverses two essential phases: meticulous data preprocessing and prediction utilizing a hybrid deep learning model. In the preliminary stage, we systematically preprocess raw data, addressing outliers, inputting missing values, and standardizing the dataset, thereby establishing a foundation of high-quality data. Subsequently, we employ a deep residual CNN integrated with an ESN architecture to proficiently model both temporal and spatial characteristics of energy consumption. Through rigorous testing on three benchmark datasets—namely the IHEPC, PJM, and AEP—encompassing both hourly and daily short-term ELF scenarios, our framework has demonstrated substantial improvements in prediction accuracy when benchmarked against existing models. This comprehensive approach not only showcases the practical viability of our model but also underscores its scientific novelty. By effectively addressing the challenges in preprocessing noisy data and capturing intricate spatiotemporal features, our framework offers a significant advancement in the field of ELF. The incorporation of diverse datasets further enhances the generalizability of our model, demonstrating its adaptability to various real-world scenarios. As we navigate the complexities of short-term ELF, our work contributes to the scientific understanding and implementation of advanced data-driven models, paving the way for more accurate and efficient electricity load forecasting.

The proposed framework for short-term ELF has demonstrated its effectiveness through rigorous evaluation on benchmark datasets. However, it is important to acknowledge certain limitations of the model. One limitation of our approach is that we used separate models for spatial and temporal feature extraction, i.e., CNN and ESN. While this design choice was effective in capturing both spatial and temporal characteristics, future work can explore more integrated architectures to streamline the forecasting process and potentially enhance the model’s performance. Furthermore, it is essential to recognize that hybrid models, such as the one presented in this study, are computationally expensive, especially when dealing with large-scale datasets. As future computational capabilities continue to advance, optimizing the computational efficiency of such models will be a priority. Another important point to note is that our current method is primarily evaluated for short-term forecasting. Expanding the model’s capabilities to handle longer-term predictions could further broaden its applicability and relevance in various energy management scenarios. In future work, we aim to address these limitations and explore the incorporation of explainable AI techniques to enhance model transparency and interpretability. Additionally, leveraging environmental sensor data and integrating forecasting with control strategies will remain essential in order to further optimize energy systems in low-carbon buildings and extend the model’s practicality. These endeavors not only aim to overcome the limitations but also advance the field of short-term energy load forecasting for more sustainable energy management across diverse temporal horizons and computational resources.