Prediction of hourly solar radiation in Indonesia using LSTM

The value of solar radiation has many advantages to support optimal utilization of solar energy, from building design, solar power plant and agricultural system. However, solar radiation is an element that will affect to other weather parameters. Forecasting systems that have been carried out by previous researchers are predictions for data per-hour, per-day, per-month and even per-year. All forecasting processes are carried out using different multivariate meteorological parameters. Fluctuations in solar radiation can occur due to data errors or moving fast of clouds. In this study, we are using meteorological parameter: solar radiation, sun duration, relative humidity, dew point, air temperature, and sky cover. All these parameters will be used to predict hourly solar radiation. Long short-term memory (LSTM) is proposed in our works to discover best activation: ReLU, sigmoid, tanh. As a result, RELU is best activation which has the smallest Mean Square Error (MSE) value is 0.1885, compared with other activation value: sigmoid MSE value is 0.2359 and tanh MSE value is 0.3270. By using ReLU, the wall time is 1min 51s and calculation process stops in 18th epoch. Optimum learning rate become 0.00249 and wall time 1min 38s. The output prediction is hourly solar radiation one-day ahead that we can used as a data solar radiation information.