Enhanced Machine-Learning Techniques for Medium-Term and Short-Term Electric-Load Forecasting in Smart Grids

Abstract

Nowadays, electric load forecasting through a data analytic approach has become one of the most active and emerging research areas. It provides future consumption patterns of electric load. Since there are large fluctuations in both electricity production and use, it is a difficult task to achieve a balance between electric load and demand. By analyzing past electric consumption records to estimate the upcoming electricity load, the issue of fluctuating behavior can be resolved. In this study, a framework for feature selection, extraction, and regression is put forward to carry out the electric load prediction. The feature selection phase uses a combination of extreme gradient boosting (XGB) and random forest (RF) to determine the significance of each feature. Redundant features in the feature extraction approach are removed by applying recursive feature elimination (RFE). We propose an enhanced support vector machine (ESVM) and an enhanced convolutional neural network (ECNN) for the regression component. Hyperparameters of both the proposed approaches are set using the random search (RS) technique. To illustrate the effectiveness of our proposed strategies, a comparison is also performed between the state-of-the-art approaches and our proposed techniques. In addition, we perform statistical analyses to prove the significance of our proposed approaches. Simulation findings illustrate that our proposed approaches ECNN and ESVM achieve higher accuracies of 98.83% and 98.7%, respectively.

1. Introduction

Electricity is an essential part of our daily lives and has a significant impact on the activities performed by individuals working in different fields. Due to the rapid population growth, there is a high demand for electricity in all parts of the globe [1]. Because of the limited capabilities of traditional electric grid stations, they are being replaced with the latest digital power grid system called smart grid (SG). Through SG, the management of electric-load distribution has become easier for utilities. It is also beneficial to minimize the discrepancy between electricity supply and demand. Controlling the generation, supply, and consumption of electricity is a considerably significant task for SG. One pathway of interaction between utilities and customers is smart meters (SM), which are part of the SG. Another important element of SG is demand side management (DSM), which is used to shift high-voltage equipment from hours of high consumption to hours of low consumption [2]. DSM is beneficial to control energy consumption according to power generation [3,4]. Figure 1 shows that load control, energy efficiency, renewable energy integration, better power quality, and plug-in hybrid electric vehicle functions are also possible using SG.

SG is also very useful in the planning of electric-load transmission and distribution. Transmission planning is necessary for utilities. It establishes the regions where electric-load expansion is needed to maintain the stability and pace of development. It also guarantees the efficient generation and distribution of electricity by the generators [5,6,7]. Similarly, distribution management is a robust method that helps meet specifications, such as determining the location, size, and installation of distribution facilities. By forecasting future load, electric-load generation and consumption can be better planned.

The forecasting of electric load helps utilities understand future demand patterns and plan according to the demand [8,9]. The utility risk is minimized by forecasting and understanding the energy patterns, which helps meet the energy demand. It helps the utilities determine the demand of the consumers and set an appropriate time-frame for maintaining the power supply in residential areas [10,11,12,13,14,15,16]. Predicting the amount of electricity needed to operate and manage the supply chain also helps the utility in avoiding excess power. In the event of increased load demand, the utility looks for a cost-effective generation while maintaining stability.

The authors in [17] state that short-term load forecasting (STLF) can forecast the electric load from one day to a week, and medium-term load forecasting (MTLF) technique is employed to anticipate electric load from one week to one year in the future. Long-term load forecasting (LTLF) is utilized if the future load forecast is required from one year to several years ahead. In this research, STLF and MTLF are performed using a large dataset. Data analysis is used to perform computational analysis of datasets.

Data analysis is a technique to extract important information from hidden data patterns. In the real world, data have very complex forms [18,19]. In recent years, the scale of real-world data has become very large, so we refer to it as big data because of its large scale. The information about electricity obtained from big data is used by utilities to perform load analysis, which can be used to create efficient and reliable electric-load management and distribution plans. In recent decades, power forecasting has been carried out using various models. However, industrialization and urbanization have increased energy consumption as a result of the depletion of natural resources. Therefore, there is an urgent need to improve the current power-forecasting frameworks and develop new ones to address these concerns and increase awareness of customers so as to make them active participants of the SG community [20,21].

The research community has presented several machine-learning (ML) and deep learning (DL) forecasting techniques to deal with the issue of electric-load forecasting [22]. However, each technique has its advantages and limitations. The primary purpose of load forecasting is to achieve reasonable accuracy rates. For large datasets, redundant features are an obstacle to more accurate results [23,24,25]. Furthermore, conventional methods are not suitable to deal with these large datasets [26] and the manual setting of hyperparameters is error-prone and increases computation time [15,27,28].

In this paper, we present a feature selection and extraction approach that excludes less important variables to solve the above problems in accurately predicting electrical load. In the regression section, an enhanced convolutional neural network (ECNN) with more layers and dynamic adjustment of hyperparameters using random search (RS) algorithm is proposed. An enhanced support vector machine (ESVM) is also proposed and RS is used to adjust its hyperparameters. Following are some of the key contributions of this study:

• Focusing on big data generated by the SG community, an effective and efficient data-analytic approach is presented for electric-load prediction to ensure grid stability.
• The parameters of the support vector machine (SVM) and convolutional neural network (CNN) are dynamically set using the RS method in our proposed ESVM and ECNN techniques for electric-load prediction.
• A feature selection and extraction model using a mix of XGB and RF approaches is proposed to address the challenges of large datasets’ computational complexity, feature selection, and feature extraction.

The rest of the manuscript is organized as follows: State-of-the-art literature review, focused research areas, and motivation of this study are discussed in Section 2. Proposed system model is described in Section 3 and results of our study are presented in Section 4 along with relevant discussion. Finally, Section 5 concludes this study and presents potential future research directions.

2. Literature Review

To ensure grid stability and to effectively manage electricity user demand, SG continuously monitors and manages the electric load with the help of efficient and effective load-forecasting algorithms. The research community has proposed different electric-load forecasting techniques to support grid stability and efficient grid management. A brief description of the current literature on electric-load forecasting is given below.

Linear regression (LR) was used by the authors of [29,30] to analyze the dependent variables and to identify the independent variables in a regression-based approach. The independent variable was considered first because it varied the most. The dependent variable in load forecasting is generally energy demand, which depends on the electricity supply. Contrarily, the independent variables are typically weather-related, such as wind, temperature, and humidity. Simulations proved the efficacy of the proposed approach.

Artificial neural networks (ANN) may be used to carry out nonlinear simulations and adjustments as we do not need to know the load and weather elements in advance. The ANN reacts as new data come into view. Currently, it is employed to address issues in power systems such as topological immobility, alarm production, fault detection, and security assessment [27,31,32].

In [33], time-series analysis (TSA) was performed on correctly consecutive data at constant intervals to obtain optimal results. This technique is used to understand the sequence of data and predict the possible outcomes depending on previous values. The proposed approach was used to predict electric load over a limited time period. Expert systems have a high degree of intelligence. As new data are fed into the expert system, it can learn more and expand its knowledge. Knowledge engineers are used in expert systems to gain knowledge and develop new prediction models for load forecasting. The authors of [34] used expert systems for electric-load forecasting.

Fuzzy logic (FL) is a logical system similar to Boolean logic [35]. In Boolean logic, 1 and 0 are used as input values. In FL, comparisons are used to generate inputs. Mathematical formulas are not used in this method to convert input values into output values. Noise does not affect the fuzzy logic. Defuzzification is proposed in [35] to obtain an accurate electric-load forecast. In [36], the authors performed predictions using different datasets of smart homes with a data-analytic approach, but could not properly manage the big data. In [37], predictions were performed using the long short-term memory (LSTM) and other ML techniques. Only LSTM performed better in their work.

The authors in [38] presented a model to forecast the future electricity consumption record using feature-selection and regression models. In [15], the prediction was performed using two datasets. However, the values of the hyper-parameters of the ML techniques were set manually in their research work. In [39], the authors used a three-step model to perform load and price forecasting. In the first step, conditional mutual information and flexible wavelet packet transform methods were used to separate the signals into various frequencies. The second phase involved implementing a nonlinear least square SVM (NLSSVM) and a multi-input multi-output model to demonstrate the association between price and load. In the last section, the method TV-SABC, which is a modified version of the artificial bee colony (ABC) optimization algorithm based on time-varying coefficients, was used to improve the NLSSVM parameters through a learning process.

A new feature-selection method was presented in [40]. The main contribution of their proposed model was to develop an interaction between the relevance and redundancy of features for the best feature selection. However, the computational time of their research was increased due to the manual tuning of the hyperparameters of ML techniques. In [23], a load-forecasting model was implemented using CNN, EPNET and LSTM. The effectiveness of the suggested strategies was assessed using the mean square error. CNN surpassed the latest available methods; however, the presence of redundant features introduced redundancy to the data and increased the time complexity. In [24], the authors used SVM as a regression technique to perform electric-load forecasting. However, SVM was not able to handle a large dataset.

The authors of [41] employed grey correlation analysis to choose the features, and used kernel function and principal component analysis to extract the features. They eliminated the redundant features of the dataset but used the conventional SVM technique to predict the load. Deep LSTM with DNN was used to predict price and electric load. These techniques improved accuracy and provided good results on large datasets. Sophisticated results were obtained by implementing a deep auto-encoder technique in [25]. However, The necessity of removing unnecessary features was not taken into account by the authors. A gated recurrent unit was presented in [42] to forecast the record of electricity consumption and price. However, the authors failed to achieve high accuracy rates using this technique. In [43], the authors proposed ESDM and DCNN to predict the electric load. The parameters of SVM were dynamically adjusted and the number of layers of CNN was increased to obtain better results. The DCNN algorithm provided better results in their proposed work.

In [44], the next-day prediction was performed to increase the layers of ANN and the optimizing algorithm. The proposed algorithm was also compared with some traditional ML techniques to show improved higher accuracy rates. In [45], a deep CNN was proposed by the authors for predicting the weekly load for the upcoming days. Nevertheless, a small dataset was used for this. A hybrid of CNN and LSTM is utilized for electric consumption record prediction in [28]. Redundancy was not taken into account, and they used data from three years. The authors of [46] used SVM and extreme learning machine (ELM), respectively. However, they manually adjusted the parameters and worked with limited datasets. For electric-load prediction, the authors of [47] combined CNN and gated recurrent unit methods. Additionally, earth-worm optimization (EWO) was used to dynamically modify the CNN–GRU hyperparameters. The suggested approach worked well; however, the authors examined electric-load data from three years only.

To improve the accuracy of the 168-hour forecasts, authors of [48] proposed a collection of ML models using historical load, weather, and holidays data. The authors did not consider removing less-important features and used conventional methods for forecasting. The primary goal of the analysis in [49] was to use a CNN-based model to incorporate traditional elements (weather, holidays, etc.) as well as current COVID-19 pandemic trends and how they relate to the STLF issue. Although the research is useful for future pandemics, it uses conventional techniques for prediction. The outcomes can be further enhanced by adjusting CNN’s hyperparameters.

The authors of [50] analyzed the ISO-NE dataset using a novel machine-learning technique, which contains daily electricity consumption data for eight years. The best features were extracted using DT and RF classifiers. The authors used SVM and CNN for the prediction of electric load and cost. Coronavirus herd immunity optimization technique was used to enhance efficiency by modifying the hyperparameters and the proposed technique was used as a classifier to enhance the performance. By adding an extra layer to the CNN and adjusting its parameters, the likelihood of an overfitting classifier is decreased. Statistical results inferred that the presented approach performed well.

To eliminate redundancy in [51], the authors employed a three-step process that includes feature selection, feature extraction, and feature prediction. The hybrid approaches XGB and DT were used for feature selection. The redundant bits of the RFE were removed using a feature extraction approach. The classification and prediction capabilities of SVM and ELM were improved using machine learning methods. GA was used to modify the hyperparameters of ELM, and the grid search algorithm improved SVM. The results show the superiority of the proposed strategy over their counterparts.

For feature importance, an XGBoost and a decision tree hybrid feature selector were proposed by the authors in [52]. The recommended structure discovered and produced a small set of fuzzy rules that focused on building electricity usage behavior using time-series data from prior operations. The responsiveness of the fuzzy system is shown and an assessment of its performance is presented, which reveals that the generated rule base has higher accuracy. Furthermore, a smaller overall set of rules is created and compared to the default decision tree configuration for assessment.

To enhance the performance of STLF, the authors in [53] presented a two-step PSO technique. Through the initial step, PSO was used to categorize the ideal input shapes for the neural network. Subsequently, the available training data was again divided into homogeneous clusters using PSO. A distinct neural network was used for each cluster. In a bus electricity consumption forecasting issue, experimental findings validated the resilience of the presented methodology, and the proposed approach was tested on a load-profiling issue, which performed better than the most popular techniques in the literature on load profiling.

The authors of [54] developed an inverse and discrete PSO method to improve the variable mode decomposition method for forecasting of next day electricity price on the basis of previously recorded weather data and utility bill information from the Greek electricity market. The forecast results were reviewed to determine if either of the two presented divide-and-conquer preprocessing approaches provided a better estimate of the short-term electric utility cost. The variational mode decomposition-based method that results in fewer mistakes in power price prediction was enhanced by the proposed variation of PSO, which had an average absolute percentage error value of 6.15%.

In [55], the authors proposed a framework for cluster-based ensemble prediction that uses an adaptive selection process to choose ensemble members for stacking and tweaking regressors. The prediction accuracy on peak and off-peak data were tested with the presented method for developing structurally adaptable estimators for each cluster. Results of the studies showed that, in this context, more reliable ensemble models were generated by associate selection techniques that focused on the effects of off-peak performance. Compared to the standalone estimators, the ensemble models performed better overall.

According to the literature review, most authors performed their predictions using conventional techniques, which have certain limitations. By improving these conventional techniques, the electric-load prediction accuracy rate can be increased. Most of the authors used small datasets, which are not very useful for predicting future loads. In this work, our main goal is to improve the accuracy rate using big data. Redundancy was also not considered in many articles. To eliminate redundancy, we used RF, XGBoost, and RFE techniques in our proposed ESVM and ECNN methods.

3. Proposed System Model

In this study, a three-stage model is proposed to accomplish STLF and MTLF. Two techniques, ECNN and ESVM, are proposed to perform accurate electric-load forecasting. The presented system model used is shown in Figure 2. The number of steps performed is given below:

3.1. Input Data

In the proposed work, eight years of huge electricity record from January 2011 to December 2018 was utilized for electric-load forecasting. These big-forecast data were downloaded from the website ISO/NE [56]. Temperature, humidity, weather, congestion, etc., are just a few examples of the independent and dependent data that make up the dataset. Our target data is the column named “SYSLoad” which represents the system load. The other features related to the target data are the day-ahead cleared demand, clearing price for the regulation market, real-time demand, dew point temperature, local day-ahead marginal price, dry bulb temperature, day-ahead energy component, real-time marginal loss component, day-ahead congestion component, real-time congestion component, and the day-ahead energy component. We used eight years or 96 months of data because the consumption patterns of similar months are roughly the same. Finally, 80% of the data are selected for training, 10% for testing, and 10% for validation.

3.2. Feature Selection and Extraction

The computational complexity of the model might be increased by several less significant features that are typically present in big datasets. If these less-important features are efficiently eliminated, then the complexity of the model can be minimized. Our proposed feature selection and extraction methods effectively eliminate the less-important features.

Statistical mechanics is applied to the dataset for the feature-selection procedure of the presented model. The importance of each feature is calculated to select the most relevant features. Accurate results are obtained by combining the XGB and RF techniques, as shown in Figure 3. A threshold is also set to exclude less-important features. The feature selection is conducted according to Equation (1).

$$f\left(s\right) =\left\{\begin{array}{c}if, XGBi\left(f\right) + RFi\left(f\right) \ge t,\hfill \\ Dropif, XGBi\left(f\right) + RFi\left(f\right) < t\hfill \end{array}\right.$$

(1)

XGBi indicates the features calculated with XGB, and RFi indicates the features calculated with RF. The features are represented by the symbol f and the threshold by the symbol t.

Feature extraction is performed after feature selection by RFE. Feature extraction selects only non-redundant features having a significant influence on the intended features of the dataset. RFE recursively collects features, uses those features to build a model, and then evaluates or reports the accuracy of the presented model. To forecast the target variable, RFE can integrate many characteristics. The following sub-sections describe feature-selection and feature-extraction techniques in detail.

3.2.1. Extreme Gradient Boosting

This technique belongs to an open-source library and is the enhanced version of the decision tree (DT), which provides more reliable and accurate results than DT. It is based on the assumption that the overall prediction error is minimized when the best feasible future model is merged with earlier models. According to this work, the importance of each feature of XGB is calculated using a scale of 0 to 1, with the more significant feature containing a value close to 1 and the less-important feature having a value close to 0.

3.2.2. Random Forest

Numerous DT methods are combined to create the RF. By bagging or bootstrap aggregation, the “forest” that the RF algorithm creates is trained. By grouping machine-learning algorithms, the bagging meta-algorithm improves their accuracy. It predicts the mean or average value of the used DT techniques in the forest. Compared to DT, RF provides more accurate results by reducing the problem of overfitting the dataset. We calculate the importance of the features using RF.

3.2.3. Recursive Feature Eliminator

This is a machine-learning technique employed for best feature selection and elimination of weak features. Through RFE, in this work, each feature is converted into a true or false dimension. Then, a threshold is set to eliminate less-important features and only the most important features are fed to the regression model.

3.3. Regression of the Load

Many authors have applied various ML and DL techniques to perform regressions. However, they still have problems with overfitting and accuracy. In our work, we improve both ML and DL methods. We prove that the problem of overfitting could be minimized, and the accuracy of load forecasting could be enhanced, if these methods are properly tuned based on the threshold value. SVM belongs to ML and CNN belongs to DL methods. RS refers to an algorithm that uses some kind of randomness or probability. Therefore, the RS is used to tune the hyperparameters of CNN and SVM and we call these techniques ECNN and ESVM, respectively.

3.3.1. Random Search Algorithm

The objective function’s random inputs are created and evaluated by the random search algorithm. This is helpful because it does not assume anything regarding how the objective function is structured. It allows for the identification of counter-intuitive solutions and can be helpful in issues when there is a lot of expertise that may impact or bias the optimization strategy.

3.3.2. Enhanced Convolutional Neural Network

The combination of RS and CNN is used to propose ECNN in this study. The CNN technique is a part of deep learning and can have several layers. The first layer of CNN is the convolutional layer, which serves just as an input filter. When the same filter is applied multiple times, a feature map is created, which gives the positions and intensity of the detected features in the given big data. The dense layer is the second layer of our proposed CNN. This layer is utilised to thoroughly link all of the neurons that were received from the layers that were connected before. After connecting all the features, overfitting may occur. To avoid this problem, the dropout layer can be used. The pooling layer is also a part of the CNN, and we have used the max-pooling layer in the proposed CNN. Max-pooling selects the most salient features from the feature map. To improve the efficiency of the CNN, we have extended the layers. Moreover, the parameters of the CNN are dynamically set using RS. The individual layers and parameters are explained in

Table 1. Layers and parameters used for proposed ECNN.

Parameters	Values
Sequential Model
Conv1D	No. of Neurons = 64
	Kernel Size = 2
	ReLU Activation-function
Dense	Parameters tuning with RS
	Units = 10
	ReLU Activation-function
Dropout	0.000001
Maxpooling1D	Size of Pool = 2
	Padding = same
Dense	Parameters tuning with RS
	Units = 50
	ReLU Activation-function
MaxPooling1D	Size of Pool = 2
	Padding = same
Dense	Parameters tuning with RS
Compiling Model
Loss Function	MSE
Metric	Accuracy
Optimizer	Adam
Training Model
Epochs	200
Verbose	0
Validation split	0.30
Batch size	10

, and the system model of the presented enhanced convolutional neural network (ECNN) is exhibited in Figure 4.

3.3.3. Enhanced Support Vector Machine

Support vector machine is a well-known supervised machine learning algorithm and mutual adjustment of its parameters is a significant challenge. In our proposed ESVM, the parameters are adjusted using the RS algorithm. As kernel, the radial basis function (RBF) is utilized. The ESVM has 15 iterations, while the C values and gamma values of the SVM are tweaked using RS. The proposed system model for the presented ESVM is shown in Figure 5.

4. Results and Discussions

The simulations are performed with Anaconda Spyder3 software. The user system has 12 GB RAM, a Core i5 processor, and belongs to the sixth generation. The Python language is used for the implementation of this research.

4.1. Feature Selection and Extraction

The importance of features is shown in Figure 6 and Figure 7, which is calculated by the feature-selection method. The importance of all features is represented on a scale from 0 to 1. Using RFE, the best features are chosen, while the worst features are removed. In this analysis, each feature is transformed into a true or false dimension using RFE. Features accepted by RFE have true dimensions and rejected features have false dimensions. Features rejected by RFE are ‘DA_LMP’, ‘DA_EC’, ‘DryBulb’, and ‘DewPnt’. The threshold for XGB is 0.8 for feature importance calculation.

The RFE = True is set to eliminate irrelevant features. After the feature selection and extraction methods, three features, DA_LMP, DryBulb, and RT_CC, are removed from the total number of features by RFE. The threshold for RF is 0.7 for feature importance calculation. The remaining features are forwarded for regression.

4.2. Regression of Electric Load

After removing the less-important features, only the data from the most important features are passed to the regression model. Figure 8 given below shows the normal load.

In our work, STLF and MTLF are performed for 1 week, 1 month, and 4 months load forecasting, as shown in Figure 9, Figure 10, and Figure 11, respectively.

The different colored lines represent these predicted values. The red line represents the actual load. Lines most similar to the actual load represent high accuracy, whereas less-similar lines represent low accuracy. In the above figures, it can be seen that the prediction line of the proposed ECNN technique is more similar to the actual load line. The proposed ESVM has a lower similarity than ECNN. However, there is a big difference between the actual load line and the lines of the conventional techniques.

4.3. Performance Evaluation

Performance evaluation is very important to assess the efficiency and effectiveness of any proposed algorithm. In this study, we have used four evaluation metrics namely: mean absolute percentage error (MAPE), root mean square error (RMSE), mean absolute error (MAE), and mean squared error (MSE) to evaluate the performance of our proposed techniques. Figure 12 and

Table 2. Error values (%).

Techniques	MSE	MAE	MAPE	RMSE
SVM	12.4	10.5	1.7	12.3
Proposed ESVM	10	9.5	1.3	9.02
CNN	3.03	6.9	14.0	2.1
Proposed ECNN	1.17	1.2	11.9	1.4

display the error rates of the proposed techniques and the most recent techniques.

The results presented in

Table 3. Accuracy rates (%).

Techniques	MSE	MAE	MAPE	RMSE
SVM	87.6	89.5	98.3	87.7
Proposed ESVM	90	91.5	98.7	90.2
CNN	96.97	93.1	86	97.9
Proposed ECNN	98.83	98.8	88.1	98.6

show a big difference in accuracy between the proposed techniques and the conventional techniques. The accuracy of load forecasting is increased by our proposed techniques. The

Table 4. Statistical analysis tests for proposed techniques and conventional techniques. Note: PSHT: Parametric statistical hypothesis tests; CT: Correlation test; NSHT: Nonparametric statistical hypothesis tests.

Techniques	Test	NSHT		CT		PSHT
		Wilcoxon Test	Kruskal Test	Pearson’s Test	Kendall’s Test	Chi-Squared Test	ANOVA Test
SVM	F-Statistics	104,549	26.0771	−0.0404	−0.036	158,449.28	30
	p-value	0.000	0.000	0.257	0.143	0.000	0.000
Proposed ESVM	F-Statistics	132,003	0.2949	−0.037	−0.0362	164,404.40	0.064
	p-value	0.805	0.588	0.317	0.144	0.000	0.803
CNN	F-Statistics	37,953	1.4537	0.996	0.949	576.09	1.3971
	p-value	0.000	0.228	0.000	0.000	1.000	0.238
Proposed ECNN	F-Statistics	131,225	0.0001	0.736	0.5321	37,915.93	0.6539
	p-value	0.655	0.990	0.000	0.000	0.000	0.418

lists numerous tests based on correlations, as well as parametric and nonparametric statistical analyses based on hypotheses for both the new methods and the traditional ones.

The results of the various statistical tests are covered in Table 4. A value of 0 indicates that the hypothesis is accepted, and a result greater than 0 indicates that it is rejected.

5. Conclusions

In this work, we studied the power-load prediction problem using an improved framework based on feature selection, extraction, and regression. The main objective of the efficient and effective electric-load prediction for big data is successfully achieved using our proposed ECNN and ESVM forecasting models. Furthermore, our proposed forecasting models helped decrease computational complexity of the forecasting model by eliminating less-important features using modern feature-selection and extraction methods. The numbers of layers of our proposed ECNN are increased and the hyperparameters of the proposed techniques ECNN and ESVM are dynamically adjusted. Simulation results of our proposed techniques are compared with conventional CNN and SVM techniques using four performance error estimators, i.e., MAE, RMSE, MAPE and MSE. The performance metrics proved that our proposed ECNN and ESVM electric-load forecasting models have the lowest error rates.

Due to the growing worldwide interest in reliable and sustainable energy supply, incorporating more renewable and alternative energy sources reduces stress on existing electric transmission systems. The proposed schemes should be helpful in finding the exact power generation from distributed sources and power consumption that helps in smooth working of the smart grid. The same infrastructure can be implemented for industrial power-management systems and will also be effective for smart agriculture systems.

A continuous network service should be required for the smooth working of SG. In a disaster situation, the smart grid faces significant performance or network congestion issues. Mobile network operators cannot guarantee adequate service during severe weather events such as storms, torrential rain, or lightning strikes. Due to vulnerabilities in the infrastructure used for implementation, smart meters could be hacked and exploited to alter electricity consumption.

In the future work, the proposed model will be further optimized using heuristic techniques, and further tests will be performed for LTLF. Renewable energy sources are also included to enhance the stability of the smart grid. We intend to perform experiments using our proposed model with another forecasting model to establish whether they offer better accuracy and convergence time. The suggested structure could be improved by expanding the current system to include a feedback module that may control the desired behavior of the residential buildings depending on particular thresholds established by the electricity provider. The proposed forecasting model will help to detect electricity thefts using classifiers. Different authentication and control access parameters should be implemented to avoid vulnerability in SG infrastructures.