Multilayer Perceptron (MLP) and Autoregressive Integrated Moving Average (ARIMA) Models in Multivariate Input Time Series Data: Solar Irradiance Forecasting

Abstract

Solar irradiance needs to estimate power consumptions for requiring of saving energy. The demand accomplished with providing facilities to predict. Time series data is a dataset that has complex problems. Using multilayer perceptron (MLP) and autoregressive integrated moving average (ARIMA) with multivariate input to solve the problem of predicting solar irradiance. The dataset is collected from solar irradiance sensor by an online monitoring station with 10 minutes data interval for 18 months. Prediction experimented with t, t-2, and t-6 data inputs that represent t as the day to get the predictive model (t+1). In ARIMA model, optimization was obtained in the input parameter (t-6) and ARIMA(1,1,2) with minimum RMSE is 43.91 W/m2, whereas MLP model used single layer, 10 neurons and using relu activation function to predict with minimum RMSE is 8.68 W/m2 using (t) input parameter. The deep learning model is better than the statistical model in this experiment. RMSE, MSE, MAE, MAPE, and R2, are used as an evaluation for model performance.