SARIMA and Artificial Neural Network Models for Forecasting Electricity Consumption of a Microgrid Based Educational Building

We develop Seasonal Autoregressive Integrated Moving Average (SARIMA) and Artificial Neural Network (ANN) models for predicting one-month and oneday ahead electricity consumption of a microgrid based educational building. The prediction models can provide forecasts up to hourly accuracy. For this objective, we use more than two million records of electricity consumption data imported from the smart meter system of a six-floor microgrid based educational building. We use the HyndmanKhandakar stepwise algorithm, which generates the (1, 0, 1)x(0, 1, 1)24 SARIMA prediction models. For the ANN prediction models, we use a thirty one-neurons input layer, a twenty-neurons hidden layer, and a single neuron output layer. The experiment results indicate that the ANN models produce more accurate and consistent predictions than the SARIMA models both in the one-month ahead and one-day ahead prediction contexts.