Short-Term Load Forecasting: Based on Hybrid CNN-LSTM Neural Network

Abstract

Accurate load forecasting greatly impacts how energy suppliers plan for real power generation, distribution, system maintenance, and electricity pricing in their operations centers. Owing to the non-linearity in most time-series datasets developing a single linear or nonlinear model is insufficient for predicting and forecasting load power consumption. This study constructs a hybrid forecasting model that combines conventional neural network (CNN) with long short-term memory network (LSTM) for short-term forecasting of time series to exploit the upper performance of a hybrid model on load forecasting over an individual LSTM model. The suggested architecture addresses both linear and nonlinear patterns in the actual dataset, allowing it to extract more precise features to better forecast power consumption. The suggested models rely on merging two deep learning architectures in a hybrid model to forecast more precisely the load consumption of a household-based in Morocco. Hence, the originality of this work resides in predicting load consumption over different time intervals by utilizing the record measurements from various former weeks. Thus, using records from the past two weeks resulted in highly accurate predictions for the one 1 day ahead and 5 days ahead of time windows with MAE of 5.48 and 6.51 respectively. However, the last three weeks' data was more valuable for the LSTM model resulting in more efficient predictions when compared to CNN-LSTM with an MAE of 6.69 and 8.45 respectively for the three days ahead of the time window and 7.59 and 8.84 for the five days ahead. The empirical results show that the proposed model adequately learns the load consumption patterns of the household and accurately predicts the short-term power consumption.

Topics & Concepts