Seasonal forecasting of the hourly electricity demand applying machine and deep learning algorithms impact analysis of different factors

Abstract

The purpose of this paper is to suggest short-term Seasonal forecasting for hourly electricity demand in the New England Control Area (ISO-NE-CA). Precision improvements are also considered when creating a model. Where the whole database is split into four seasons based on demand patterns. This article's integrated model is built on techniques for machine and deep learning methods: Adaptive Neural-based Fuzzy Inference System, Long Short-Term Memory, Gated Recurrent Units, and Artificial Neural Networks. The linear relationship between temperature and electricity consumption makes the relationship noteworthy. Comparing the temperature effect in a working day and a temperature effect on a weekend day where at night, the marginal effects of temperature on the demand in a working day for power are likewise at their highest. However, there are significant effects of temperature on the demand for a holiday, even a weekend or special holiday. Two scenarios are used to get the results by using machine and deep learning techniques in four seasons. The first scenario is to forecast a working day, and the second scenario is to forecast a holiday (weekend or special holiday) under the effect of the temperature in each of the four seasons and the cost of electricity. To clarify the four techniques' performance and effectiveness, the results were compared using the Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Normalized Root Mean Squared Error (NRMSE), and Mean Absolute Percentage Error (MAPE) values. The forecasting model shows that the four highlighted algorithms perform well with minimal inaccuracy. Where the highest and the lowest accuracy for the first scenario are (99.90%) in the winter by simulating an Adaptive Neural-based Fuzzy Inference System and (70.20%) in the autumn by simulating Artificial Neural Network. For the second scenario, the highest and the lowest accuracy are (96.50%) in the autumn by simulating Adaptive Neural-based Fuzzy Inference System and (68.40%) in the spring by simulating Long Short-Term Memory. In addition, the highest and the lowest values of Mean Absolute Error (MAE) for the first scenario are (46.6514, and 24.759 MWh) in the spring, and the summer by simulating Artificial Neural Networks. The highest and the lowest values of Mean Absolute Error (MAE) for the second scenario are (190.880, and 45.945 MWh) in the winter, and the autumn by simulating Long Short-Term Memory, and Adaptive Neural-based Fuzzy Inference System.

Keywords: Adaptive neuro-fuzzy inference system; Day-Type; Electricity price; Gated recurrent units; Hourly demand forecasting; Long short-term memory; Mean absolute percentage error; Temperature; Time-series forecasting.

Fig. 1

A structure model of the LSTM network.

Fig. 2

A structure of the GRU.

Fig. 3

The fundamental architecture of ANFIS for a single output with two rules and two inputs.

Fig. 4

Predicting the hourly energy load framework using the two scenarios’ performance calculations and the four suggested methodologies.

Fig. 5

The Sys_demand (MWh) hourly dataset for the four seasons, linked with the associated hourly Market_price ($/h).

Fig. 6

Data from statistics on hourly Sys_demand and the corresponding Market_price throughout the four seasons.

Fig. 7

The actual and the forecasted results in the 1st. Scenario (workdays) of the Sys_demand (MWh), throughout a 24-h period in each of the four seasons by the four featured algorithms.

Fig. 8

The calculated performance (RMSE, NRMSE, MAE, and MAPE) for the four specified algorithms in the four seasons for the Sys_demand (MWh), 1st. Scenario (workdays).

Fig. 9

The calculated performance (The Nash–Sutcliffe Efficiency (NSE), the determination factor (R2), kurtosis coefficient, and coefficient of variation (CV)) for the four specified algorithms in the four seasons for the Sys_demand (MWh), 1st. Scenario (workdays).

Fig. 10

The actual and the forecasted results in the 2nd scenario (special holidays) of the Sys_demand (MWh), throughout a 24-h period in each of the four seasons by only two featured algorithms (GRU, and ANFIS).

Fig. 11

The calculated performance (RMSE, NRMSE, MAE, and MAPE) for the specified algorithms (LSTM, and ANFIS) in the four seasons for the Sys_demand (MWh), 2nd. Scenario (special holidays).

Fig. 12

The calculated performance (NSE, R2, kurtosis coefficient, and coefficient of variation) for the specified algorithms (LSTM, and ANFIS) in the four seasons for the Sys_demand (MWh), 2nd. Scenario (special holidays).