A study on the estimation of global horizontal irradiance via deep learning technique

For grid-connected photovoltaic (PV) systems, accurate solar irradiance forecasting is crucial, especially in cases of intermittent environmental conditions, in order to ensure grid operation, scheduling, and grid energy management. Solar irradiance, temperature, and other meteorological factors significantly influence PV generation and also make it intermittent in nature. Time series analysis provides an efficient way to map temporal patterns, like hourly, monthly, yearly, and seasonal variations, and within solar irradiance data. In this paper, time series forecasting of 24hourly future observations of global horizontal irradiance (GHI) is done using a multilayer perceptron (MLP) of deep learning framework. Different hyperparameters like hidden layers, hidden units of each layer, activation function, optimizer, and lag observations/steps are tuned to develop the best model that fits the GHI historical data. Further dropout regularization is used to avoid overfitting the model. The model is validated using validation data that the model has never seen before. RMSE, MSE, MAE, and NRMSE errors, aiming at minimizing forecasting errors on test data, are used to select the optimal value of hyperparameters.