Electrical load forecasting using adaptive neuro-fuzzy inference system

The importance of a precise load forecasting in energy system designing has been known since many years ago. To achieve an efficient, secure and economically optimal operation in an energy systems, it is necessary to analyze the power system operation in time and perform fast reactions to changes in electrical load. These are highly dependent on an accurate short-term load forecasting. One the other hand, the load prediction is also of great importance for energy suppliers, independent system operators, financial institutes and other parties involved in power generation, transmission, distribution and market. For the energy suppliers, a timely implementation of long-term load forecasting helps the system to improve the network stability and reduce the equipment failures and power outages and ensures a reliable energy supply. This dissertation have tried to propose a new approach toward the problem of load forecasting to provide more accurate results with a lesser error rate compared to the current methods. For this, a methodology based on the Chaos and Concept Drift theories is introduced, in which the Artificial Neuro Fuzzy Inference System (ANFIS) plays the main role in training and testing the model. After a short introduction on the chaos theory and nonlinear systems in the second chapter, the existence of chaos in the electrical load of the Clausthaler Umwelttechnik Forschungszentrum (CUTEC) in the year 2014 as the sample data is examined by three different methods. This chapter also deals with the concept drift as a kind of anomaly existence in time series, which can help to classify the input data based on the similarities and prepare them for a systematic feeding in the training process. The third chapter will briefly provide the basic information on electricity load prediction in general, introducing the classic and modern approaches toward the problem of load forecasting. Afterwards, the most common modern method, namely Artificial Neural Network (ANN) will be studied and the result of prediction based on a simple ANN will be presented. The fourth chapter describes a tool for analyzing the use of contextual information by fuzzy neural network. Here, the ANFIS model will be introduced and the results of weekly, monthly and seasonal electrical load prediction based on this model will be presented. The results show, that the Mean Absolute Percentage Error (as one of the main indices in load forecasting) in the ANFIS- based method is roughly 2,1% to 2,6 % which is the lowest reported rate in the similar studies with almost the same data sets and prediction horizon. Subsequently, a chapter will introduce a simulation model in MATrix Labratory (MATLAB) for a decentralized energy system with the case study of the Energy Park at CUTEC institute. This model includes different units such as combined heat and power generation units (CHP), Photovoltaics, Solar thermal panels, boiler and storage systems. The implemented model is validated with EnergyPro, a commercial energy analysis software, and the results from both models have been compared. Chapter six is dedicated to introducing the current energy system in Iran by providing information on the future development plans in renewable energies there and the transferability of the proposed model to Iran’s network system. Finally a chapter will review the methodology and results presented in this dissertation and an outline for the possible future works will be discussed.