Electricity Consumption Prediction Using Artificial Neural Networks and Evolutionary Algorithms: A Case Study Tulkarm City رسالة ماجستير

Abstract

Electricity consumption is one of the main concerns in any country; as without planning the process of the electricity distribution, there will be major problems in the economy and all the fields that affect daily life. Electric power demand prediction is a systematic procedure that allows the quantitative definition of future demand and, as it is of vital importance depending on the period that is adopted in the analysis. This study presents a model based on a supervised learning approach to predict future electricity consumption. Thus, can help power companies to plan the future demand and to guarantee the service continuation. For this purpose, real data has been collected periodically (monthly, seasonal and yearly) from Tulkarm municipality through the years 2018 to 2020 forming a sample of data points. In this thesis, we applied the collected datasets on different ANNs approaches from the traditional ones to hybrid novel models. Multilayer Perceptron Neural Networks model (MLPNNs), RNNs, and NARX have been selected from the traditional approaches, along with genetic algorithms integrated with K mean clustering for producing specific initial population seeding techniques and optimizing crossover operators to enhance the efficiency and find the optimal solution. The results showed that RNNs outperformed the other models in terms of short term memory (STM) with Root Mean Square Error (RMSE value = 5.82*10^e-24, while the hybrid models (Recurrent Neural Networks with Optimized Algorithm (RNNs-OA), Nonlinear Autoregressive Exogenous with Optimized Algorithm (NARXOA), and Nonlinear Autoregressive Exogenous with Optimized Algorithm with K Mean Clustering (NARX-OA-K Mean cluster)) gave better results, for example when executing 8 neurons for one-year prediction, RNN-OA recorded RMSE value = 0.28128, NARX-OA recorded RMSE value = 0.10382 and NARX-OA-K Mean cluster recorded RMSE value = 0.08759 which perform a good balance with the lowest RMSE especially in the long term forecasting and also outperforms other hybrid prediction models.