Long Short Term Memories for Microhydro Power Forecasting

The microhydro power plant harnesses river flow for renewable electricity production, contributing to the grid managed by the State Electricity Company (PLN). However, due to natural variables like rainfall, wind, and sunlight, accurate power forecasting is crucial for plant scheduling and daily operations. This study creates a forecasting model using Long Short-Term Memory (LSTM) with historical data consist of power, water level, and rainfall data in which simulation employs optimally tuned parameters with comparison to Artificial Neural Network (ANN) model. Results over day-ahead prediction shows that LSTM performance is shown as 79.05 and 0.12% for RMSE and MAPE, respectively. For ANN, the RMSE and MAPE are 46.60 and 0.12%, respectively. Meanwhile, results over week-ahead test period show LSTM with RMSE of 84.82 and MAPE of 0.01%, while ANN yields RMSE of 78.95 and MAPE of 0.01%. The LSTM model takes 52.6 seconds for 168-hour forecasting, compared to ANN's 0.27 seconds.