Short-term load forecasting of Gothatar feeder of Nepal Electricity Authority using Recurrent Neural Network

Abstract

This paper mainly focuses on short-term forecasting, gives an hourly demand forecast of electricity. Forecasting using Recurrent Neural Network (RNN) is helpful for making important decisions in the field of preventing misbalancing in load and power generation, scheduling, load switching strategies, preventing imbalance in the load demand and power generation, thus leading to greater power quality and network reliability. We use a method called Recurrent Neural Network (RNN) to anticipate the future hourly demand of Gothatar feeder, Nepal Electricity Authority (NEA). The RNN Network is build, trained and test with historical hourly demand data along with six different input variables and used for the prediction of day ahead hourly demand. The output from RNN model is validated with the real hourly demand data collected from NEA. In addition, the load forecasting is performed for short term load forecasting (STLF) using some other time series methods like: Single Exponential Smoothing (SES), Double Exponential Smoothing (DES) and Holt-Winter's method as well, and whose output was compared with that of RNN. The Root Mean Square Error (RMSE) of SES, DES and Holt-Winter's method was found to be 188.033 kVA, 181.066 kVA and 169.759 kVA respectively, degree of Determination (R 2 ) was 0.609, 0.618 and 0.634 and Mean Average Percentage Error (MAPE) was found to be 15.421%, 13.31% and 11.502% respectively. The RNN method proved to be the accurate and best forecasting method when the results are compared with other forecasting methods in terms of different error measurements i.e., RMSE, R 2 and MAPE. Root Mean Square Error (RMSE) of 69.03 kVA, R 2 of 0.876 and MAPE of 4.35% obtained from RNN. So, the RNN model proved to be the most accurate and best method with very less error and better R2 in this study