Multi-Variate, Recurrent Neural Network in a Short-Term Time-Series Substation Demand Forecasting

Abstract

Aggressive increase in demand in Saudi Arabia is a major concern for National Grid Network planning engineers for over a decade. Using sophisticated commercial software such as SPSS, SAS and even excel-based forecasting had been delivering results by planning engineers preparing for the future of the kingdom. Neural Network has been so powerful in today’s digital transformation, and it is known as useful in forecasting. This paper demonstrates and uses a different Neural Network structure called Recurrent Neural Network (RNN) the Long-Short Term memory (LSTM), to capture and predict substation demand behavior. Temperature, temperature dewpoint, and historical demand are the features used to predict the short-term demand of high-voltage substations located in Jeddah. A high-dimensional, preprocessed with a year-long hourly historical substation demand data is utilized. Using a sophisticated anomaly detection algorithm, Isolation Forest to track outliers of the preprocessed data. The MSE result of preprocessed and sanitized significantly reduced from 4.257 to 3.959 respectively. RNN-LSTM structure has a week-long (168 data points) timesteps with 3 input layers or features, 3 hidden layer neurons coupled with 20% dropouts in each layer densely connected to produce a month-long demand forecast. Consideration for the selection of activation functions would also ease the requirement of computing time which is reduced with an average of 5 seconds per epoch in this model when using RELU activation function.