Load Forecasting on 150 kV Grid Substation Based on Spatiotemporal Deep Neural Network

Abstract

A smart grid concept with a prediction system provides accurate information as an early warning in the process of generation and distribution of electrical energy. The complexity of the power distribution system that involves complex parameter settings is a major challenge that is difficult to predict. Although the data recording in the smart grid system is done centrally, however, the prediction system faces a lack of accuracy due to incomplete records and low data quantity. In this paper, a spatiotemporal deep neural network is developed based on convolutional long short-term memory (Conv-LSTM) to extract long-term short-term patterns in an electrical load dataset. This data set is obtained from daily measurements for six months at Cawang Baru Substation, Indonesia. The proposed model adopts the basic concept of multi-layer perceptron to record temporal patterns in several stages, thereby producing more accurate results. This model uses supervised learning techniques to propagate sequential data into the target which is the next event of the data series. Furthermore, the proposed architecture supports multivariate feature extraction so as to capture important correlations between multi-dimensional features. This study also uses the basic Multivariate LSTM (MV-LSTM) model and naive Machine Learning models including Logistic Regression (LR), Random Forest (RF), and Support Vector Regression (SVR) as benchmarking methods for the proposed model. In the testing process, Conv-LSTM achieves higher accuracy than MV-LSTM, SVR, LR, and RF with scores of 0.3688, 0.3645, 0.1332, 0.1438, and 0.1234, respectively, evaluated using R-squared. Finally, experimental results support the view that combining multivariate data and a spatiotemporal prediction model is superior for time series prediction tasks rather than univariate data.