Artificial Neural Networks for Predicting Building Electricity Consumption Based on Weather Conditions and Electricity Prices

Abstract

This thesis investigates the potential of using artificial neural networks (ANNs) to predict electricity consumption in buildings based on weather conditions and the price of electricity. The study deploys two different feedforward neural network models, using Levenberg-Marquardt and Quasi-Newton backpropagation algorithms. The ANNs are based on fully connected neural networks ranging in different depth and wideness.

The primary objectives are: (1) to develop and compare the two predictive models, (2) to analyse the predictions made using different models, architectures, and data processing techniques, and (3) to conduct a sensitivity analysis to determine the influence of the various weather conditions and electricity prices used as input.

Real world data from The Norwegian Maritime Museum is used as the case study for model training and validation of the ANNs. The results indicate the ANN models showing promising results in predicting electricity consumption, with notable differences in performance based on the chosen backpropagation algorithm, layer structure, and data processing. Both the models demonstrate a degree of precision but lack accuracy, meaning they follow the correct consumption pattern while failing to predict the correct numerical values consistently. Overall best prediction performance was given by the LM model. The conducted sensitivity analysis reveals that solar irradiance is the most important predictor for non-occupied hours, while ambient temperature, electricity prices, and humidity gain a more significant role during occupied hours. The analysis underscores the importance of context-specific models.

The findings suggest that integrating weather data and price in prediction models can enhance the efficiency in management of electricity usage in buildings. Future research should explore additional factors and refine model architecture to further optimize the accuracy of predictions.