Solar Irradiance Forecasting Using Neural Networks

Abstract

Accurate solar irradiance forecasting is essential for minimizing operational costs of solar photovoltaic (PV) generation as it is commonly used to predict the power output. This thesis presents and compares three different machine learning approaches of solar irradiance forecasting: Random Forest (RF), Feedforward Neural Networks (FNNs) and Long Short-Term Memory (LSTM) networks. Each model was tested on two different forecasts: the next hour average and the hourly day-ahead averages. The machine learning algorithms were trained and tested on data from a weather station located at Tampere University (TAU) in Tampere, Finland. Data were preprocessed before training the algorithms and the relevant features were selected. Moreover, Grid Search and Random Search techniques were used along with multiple train and validation splits to find the optimal hyperparameters for each machine learning algorithm. Persistence model is set as a baseline model for comparison while RMSE and MAE are used to quantify the prediction error. For the next hour forecast, LSTM achieved the highest accuracy in terms of RMSE (76.14 W/m2), 2.1% and 1.1% better than RF and FNN respectively. Instead, FNN generally produced the best results in the day-ahead forecast. In all models, the prediction error increases as the forecast horizon increases until it stabilizes at 10 hours approximately. Further, the error keeps increasing but slower. Besides, the next hour forecast models were able to predict considerably better the next hour solar irradiance than the day-ahead forecast models.