Load forecasting using ann techniques

This thesis describes a new method for the electric load forecasting using artificialneural networks. The work has been done for both the short term and the long term loadforecasting. A full study has been done on the various variables influencing the electricload. A detailed study has been done on the weather variables and their effect onindividuals and the electric demand. A novel feature classification technique has beenused for the analysis of the various inputs of the network. The system uses a feedforwardmulti-layer perceptron network that is trained using the backpropagation technique.F or the short term load forecasting the system will be used following two paradigms,one for the prediction of the weekdays and the other for the prediction of the weekends.The system produces each hour the forecast of the following twenty-four hours and usesthe last load value for its forecast. The weather variables are fully implemented in thesystem. The system is only retrained weekly with the change from weekdays to weekend.The retraining time is well below the time limit. The adaptivity, retraining frequency andtime horizon emphasizes the applicability of the system.For the long-term load forecasting, two models are used. One model to forecast themaximum demand and the other to forecast the energy consumption. The systemforecasts the following ten years. The economic and demographic indices are fullyimplemented in the system.The data used in this work is real data obtained from a major electric utility. The datacovers a large period of time. It has been used without any pre-processing. The systemhas been tested extensively. The tests have been done on various weather patterns inorder to make sure of its ability to perform for any characteristics. The results have beenanalyzed not only depending on the accuracy but also on their sustainability and thesystem requirements to produce them. These requirements, such as data availability, timeand computational requirements are realistic and simple.The use of neural networks for the electric load forecasting is shown to be superior tothe conventional methods. The obtained accuracy is better than that of the conventionalmethods. It also does not suffer from high computational requirements or numericalinstabilities. Furthermore, it combines the capability to include various variables withoutaffecting the non-stationarity characteristic of the electric load.This work presents a system for the electric load forecasting. Its performance makes ita significant step forward towards achieving a complete and accurate knowledge aboutthe future load. This knowledge will certainly improve the operational and planningprocedures of the power system. The proposed system can be used by any utility for itsload forecasting operation. It is suitable for on-line operation without the need of humanexperts.