Peak Load Forecast using Long Short Term Memory Networks.

Abstract

Forecasting is the technique of making scientific predictions of the future based on historical data trends. With increasing smart grid technologies and penetration of intermittent renewable energy technologies, peak load forecasting has been an important task in recent years for optimizing power grid operations. In recent years, machine learning has been one of the popular methods for load forecasting. In this paper, a deep learning single Long Short TermMemory (LSTM) based model is proposed for predicting the peak load demand of Nepal and compared with other statistical models. Comparing the evaluation metrics, it is deduced that the proposed LSTM model with 32 LSTM unit and a lookback of 30 has better forecast accuracy with Mean Absolute Error (MAE) of 34.01 MW, Mean Squared Error(MSE) of 2880.75MW, Root Mean Squared Error(RMSE) of 53.67MW, Mean Absolute Percent Error(MAPE) of 2.950% and 𝑅2 Score of 0.933. Among the statistical models considered, theWeighted Moving Average with a loopback of 30 days had the least forecast errors with a Mean Absolute Error(MAE) of 34.77 MW, Mean Squared Error(MSE) of 3271.68 MW, Root Mean Squared Error(RMSE) of 57.19 MW, Mean Absolute Percent Error(MAPE) of 3.005% and 𝑅2 Score of 0.921. 3