Short Term Solar Power and Temperature Forecast Using Recurrent Neural Networks

Solar energy is one of the world's clean and renewable source of energy and it is an alternative power with the ability to serve a greater proportion of rising demand needs. The operation and maintenance of solar energy have a significant impact on PV integrated distribution grids. Hence, the short-term forecasting of solar power is an important task for the effective management of grid-connected PV. In recent developments, most of the electric appliances (air conditioners, geysers, clothes dryers, electric blankets, etc.) usage mainly depends on the weather temperature. Therefore, temperature variations are considered to have a significant impact on the use of electrical appliances. Rapid solar integration and advanced temperature-dependent electrical appliances have drawn attention to the prediction of solar power and temperature in advance for efficient grid operation. Therefore, this paper proposes a Long Short Term Memory (LSTM) based forecast model for accurate forecasting. The suitable network structure for accurate forecasting of solar power and temperature is obtained by doing statistical analysis on the various network models. The statistical analysis gives the two-layer LSTM structure (i.e. layer 1 with 10 nodes and layer 2 with 20 nodes) is the suitable architecture for accurate forecasting of solar and temperature data. The proposed LSTM structure gives 0.2478 Mean Absolute Percentage Error (MAPE) and 6.7207 Root Mean Square Error (RMSE) for solar data, while for temperature data, it gives 0.014 MAPE and 1.0423 RMSE. The proposed network model showed an improvement in the forecast accuracy over the traditional network models available in the literature.

Automated summary

This paper proposes an accurate forecasting model for solar energy and temperature based on LSTM, obtaining a suitable network structure through statistical analysis. The model shows higher prediction accuracy for solar and temperature data compared to traditional network models available in the literature.