1. Introduction
In 2019, the UK amended the previous Climate Change Act of 2008 and set a more ambitious target to achieve net-zero carbon emissions by 2050 [
1]. Correspondingly, the electricity market in the UK has seen significant reforms to accommodate more power generation from renewable energy sources for emission reductions [
2].
Figure 1 shows the percentage change in power generation from different energy sources in the UK since 1998 [
3]. It can be found from
Figure 1 that the share of renewable energy generation has significantly increased since 2010, which threatens the stability of the grid and makes it more challenging to maintain a balance between generation and demand. It is well known that power generation must be equal to the load demand. If the gap in between is over a certain range, action must be quickly taken to reduce the gap to an allowable margin within a few or a few ten seconds; otherwise, the power grid stability may not be maintained, and blackout may be triggered [
4]. The high penetration of unpredictable power generation from intermittent renewable energy sources makes grid balance maintenance very challenging and costly.
The challenges are escalating with the integration of new electric devices and new types of usages due to future electrification in heating systems and transportation [
5]. According to the UK National Grid electricity system operator, the cost of balancing the UK grid reached £2.65 billion in 2021, up 48% year-on-year [
6]. The increased grid balancing costs cause a significant increase in electricity prices, while depriving renewable energy of its potential economic advantages. To minimise the gap between the power generation and usage, effective load prediction is essential, which will allow the operator to plan ahead of scheduling the power generation.
Through predicting future electricity demand, energy companies and utilities can better accordingly plan their generation capacity. This helps minimise the risk of power outages due to overloading the grid or high excess power generation over the load demand. Load forecasting also helps smooth pricing so that electricity can be provided at a competitive rate. It allows energy providers to optimise their operations by managing peak load times more effectively through better scheduling, maintenance plans and investment in infrastructure improvements. The historical load demand used for forecasting is the time series data [
7]. Time series data is a set of values that are sequentially listed in time order. Time series data often contain trends and seasonal fluctuations, which need to be considered when analysing the data. Time series data analysis has been popular in recent years and has been applied in areas such as stock price prediction and power management [
8]. The functional modelling approach is used for predicting electricity demand and electricity price, which produces superior forecasting results. Results have shown that functional modelling performs better than non-functional techniques, such as autoregressive (AR) [
9,
10]. The short-term load forecasting problem is addressed through an ensemble learning scheme. The prediction results produced by three base learning algorithms used by a top-level method are more accurate compared to state-of-the-art techniques available [
11]. Two customised ARIMA (p,D,q) were used to predict stock prices using Netflix stock historical data over five years. ARIMA (1,1,33) achieved accurate results, which shows the potential for stock forecasting [
12].
Figure 2 demonstrates the popular methods used in the last two decades [
13].
An artificial neural network (ANN) is a type of artificial intelligent algorithm that mimics the behaviour of a biological neuron. They are composed of interconnected neurons, which work together to process information and solve problems. ANNs can be used for time series prediction in a variety of ways. One approach is to use a recurrent neural network (RNN), which is an ANN that uses a sequence of inputs to make predictions. Another approach is to use a multilayer perceptron (MLP), which is an ANN that can be used to analyse time series data and make predictions. LSTM is a type of RNN that has become increasingly popular in recent years due to its ability to effectively process sequential data. It was first introduced by Hochreiter and Schmidhuber in 1997 [
14]. A total of 861 NVDI images in two selected regions were used for making the time series data, which was then used for the future vegetation dynamics forecast with LSTM [
15]. The previous top-down and bottom-up long-term load forecast methods were unable to incorporate different levels of information. Dong proposed a hybrid LSTM method using sequence prediction for this classic and important task [
16]. Long and short-term memory (LSTM), gated recurrent networks and convolutional neural networks were trained to predict daily electricity loads from one to seven days in advance [
17]. The prediction results on the test set showed that the LSTM achieved the best performance. Aragón introduces load prediction as continuous input for optimisation models within an optimisation framework for short-term control of complex energy systems, and the LSTM model is used as it allows incremental training in an application with continuous real-time data [
18]. Accurate and effective short-term power load forecasting is very important for power systems. Considering the temporal and non-linear characteristics of power load study the application of standard LSTM network and its two typical variants, the Gated Recurrent Unit and the Just Another Networking short-term power load forecasting [
19]. A multi-scale CNN-LSTM hybrid neural network short-term load forecasting model considering real-time electricity prices was proposed, achieving an accuracy of 98.78% [
20]. The prediction results provided a new way for the development of power load forecasting. To overcome the limitations of previous studies and further strengthen prediction performance, a novel short-term power load prediction system, VMD-BEGA-LSTM (VLG), integrating a data pretreatment strategy, advanced optimisation technique and deep learning structure, was developed [
21]. In order to improve the accuracy of short-term load forecasting of power systems, a combination model based on LSTM and light gradient boosting machine (LightGBM) was proposed. The experiment first decomposed historical load data by empirical mode decomposition, used historical weather data and load data decomposed by empirical mode decomposition to establish LSTM prediction model and LightGBM prediction model, respectively, and then these two predicted values were linearly combined to obtain the final predicted value [
22]. An integrated evolutionary deep learning method based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), an improved grasshopper optimisation algorithm and LSTM networks was proposed by Hu [
23]. Experimental results showed that the integrated evolutionary deep learning method proposed in Hu’s paper was an effective tool for STLF. Alsharekh developed an innovative framework for short-term electricity load forecasting, which consists of two phases: data cleaning and a residual convolutional neural network with a multilayer LSTM architecture [
24].
Support vector machine (SVM) was developed by Vapnik and Alexey in 1963 and is a supervised learning algorithm that uses labelled training data to identify the best hyperplane, which can act as a decision boundary for categorising new data points [
25]. This hyperplane partitions the input space into two or more classes by maximising the margin between them. SVM can be used in both binary and multi-class classification problems, as well as regression problems where the output is real values. The SVM is combined with the fuzzy set theory to develop a novel fuzzy twin support vector machine model, which is applied to predict stock price trends. The fuzzy twin SVM performs better when facing outliers, accounting for its better performance when handling data containing noise [
26]. M. Shao developed an SVM energy consumption prediction model to predict the energy consumption of hotel buildings, and the MSE value of the prediction result was 2.22%, and the R2 value was 0.94 after optimisation [
27]. Francis proposed ε-Descending Support Vector Machines (ε-DSVMs) to solve the non-stationary input data and reduce the number of support vectors compared with conventional SVM [
28]. A two-phases training SVM was introduced by Zhiwang [
29]. The two phases correspond to two different linear programming models that improve prediction accuracy and performance. Particle swarm optimisation (PSO) is integrated into the support vector machine to predict the thermal load, and the PSO is utilised to find the optimal SVM parameters [
30].
SVM and LSTM have been evaluated as effective time series load forecasting methods. SVM has the advantage of handling non-linear data patterns and robustness against outliers, which enables it to deal with high-dimensional datasets and requires less computational resources than other methods. LSTM can effectively capture long-term dependencies and remember information over extended periods, making it ideal for load forecasting applications. Additionally, LSTM networks are able to quickly and accurately handle large amounts of data without sacrificing accuracy. As such, they are often preferred over traditional models when dealing with high-dimensional datasets that contain complex patterns and temporal dynamics. Therefore, in this paper, two intelligent algorithms, LSTM and SVM, are applied for load forecasting; the University of Warwick campus energy data is used for algorithm refinement and verification. The novelty is to compare and analyse the prediction accuracy of two intelligent algorithms with multiple time scales and to explore better scenarios for their prediction applications. The high-resolution load forecasting over a long range of time is conducted in this paper. It is not common to predict load demand data for 48 × 7 and 48 × 30 sizes at one time. The results confirm the algorithms’ effectiveness and their suitability for different time horizon load prediction tasks.
The rest of the article is organised as follows.
Section 2 introduces the principles of the two intelligent algorithms and their development for load forecasting. Some pre-prediction work and a prediction flow chart of the two algorithms are shown in
Section 3. Prediction results of the two algorithms with multiple time scales in various application scenarios are demonstrated and compared in
Section 4. This chapter also presents a blind test. Finally,
Section 5 contains the conclusion.
5. Conclusions
In this paper, two computational algorithms, LSTM and SVM, were introduced and developed in the establishment of the load prediction model. They were adapted to conduct the load prediction based on selected features and historical campus half-hour recorded load demand data in four different prediction time horizons. Prediction performance was evaluated through two aspects: training/epoch time and prediction accuracy for comparison. It was found that both methods had higher prediction accuracy for shorter time horizon prediction scenarios, indicating better applicability in shorter time horizon prediction. Comparison showed that the LSTM had a higher prediction accuracy according to ultra-short and short-term load prediction, while SVM predicted better for medium- and long-term scenarios; the prediction accuracy was also verified through the blind test. Furthermore, the training time for SVM was shorter in medium- and long-term cases, confirming the high dependence of LSTM prediction accuracy on the computing power of the computer.
The existing platform limited the performance of LSTM in longer time horizon scenarios. So, a more complex LSTM model could be constructed in further stages with more nodes and hidden layers to improve its prediction accuracy when facing larger data sizes. The strategy of selecting hyperparameters can be improved with some algorithms, such as combining the model with the fish swarm algorithm. In addition, some more advanced artificial intelligence algorithms, which have been applied in dealing with time series data, can be compared with the existing two algorithms in more scenarios.