Development of a hydropower forecasting model at Isimba HPP on River Nile

Abstract

This study compared the performance of Multi-Layer Perceptron Feed Forward Neural Network (MLP-FFNN), Grey model, Convolutional Neural Network (CNN) and the Random Forest (RF) regressor for development of a hydropower prediction model for the 183 MW Isimba Hydropower Plant. In addition, a physics based model was obtained from the standard hydropower generation potential equations was implemented as a baseline for comparison. Prediction models based on hourly data were developed; the physics based model was used as a baseline for comparison, the RF, MLP-FNNN and CNN models for prediction of output power generation. The physics based model performed the well with NSE of 0.68 and very minimal errors MAE of 10.711 and RMSE of 14.6161. The Grey Model showed moderate improvement NSE of 0.3145, MAE of 14.6161. Among machine learning methods, the MLP-FFNN achieved the best performance with NSE of 0.752 and the lowest errors MAE of 8.72, RMSE of 13.53. Random Forest also gave good results NSE of 0.5997, RMSE of 11.713 while CNN attained the NSE 0.731 but higher errors MAE of 8.3659, RMSE of 11.6835. These results confirm that the MLP-FFNN outperformed both RF and CNN, making it the most accurate model. The supervised deep-learning model developed for the prediction of hydropower generation will have the potential to reduce on the operational and maintenance costs and increase or optimize the energy output of hydropower generation. The results can thus help policymakers and organizations to plan energy management using evidence-based forecasts and manage water and energy resources more efficiently.