Cuckoo search based adaptive neuro-fuzzy inference system for short-term load forecasting

Abstract

Short-Term Load Forecasting (STLF) is an essential input for power system operation computations to achieve proper system balancing. General economy and security of power system depend on accurate STLF. The accuracy of forecasting model depends on the number and types of the forecasting variables. Furthermore, a day-ahead hourly-load forecast has to reach the decision makers before the elapse of a pre-set time. Conventional methods used in determining future load demand were not able to explore all the available variables in a particular forecasting region. Moreover, artificial intelligence methods like Adaptive Neuro-Fuzzy Inference System (ANFIS), are associated with computational difficulties, thus influence the speed and accuracy of the model. Therefore, these variables need to be investigated so as to make the forecasting model simple and easy to use. Similarly, the forecasting speed needs to be improved. This thesis presents the development of short-term electric load demand forecasting algorithm, with the aim to improve the forecasting accuracy and speed. It starts with the development of data selection and data processing framework, through the use of correlation analysis, hypothesis test and wavelet transform. Variables of the four seasons in one year were investigated and were classified based on the available weather and historical load data in each season. To reduce the variability in the forecasting data, wavelet transform is used. The whole forecasting algorithm has been developed by integrating Cuckoo Search (CS) algorithm with ANFIS to produce CS-ANFIS model. CS was used to improve the forecasting capability and speed of the traditional ANFIS, by replacing the derivative-based gradient descent optimization algorithm with CS in backward pass. Its purpose is to update the antecedent parameters of the traditional ANFIS, through the determination of an optimal value of the error merging between the ANFIS output and targeted output. The whole system is validated by the comparison with an existing ANFIS model, and two other ANFIS models optimized with Particle Swarm Optimization (PSO-ANFIS) and Genetic Algorithm (GA-ANFIS). The developed CS-ANFIS model proved to be superior to these models in terms of accuracy and forecasting time. A reduction in average mean absolute percentage error of 84.48% for one year forecast is recorded using the developed CS-ANFIS, and 77.32% with the proposed data selection approach. The model was found to forecast the future load demand within an average period of 37 seconds, as compared to the traditional ANFIS which recorded an average forecasting time of 219 seconds. It can therefore, be accepted as a tool for forecasting future energy demand at utility level to improve the reliability and economic operation of the utility.