Study And Implementation Of Power Load Forecasting Based On Neural Networks

The electric power industry is the foundation of national industry system. The stable working of electric power industry constantly is related to the lifeline of national economy. Power peak load forecasting is an important work in power management system. Accurate power load forecasting is the basis of stable working of power system, economic and reasonable power price and real time scheduling. Especially in economic field, precise power prediction is of great importance in proper resource scheduling, optimizing planning of power production, achieving optimal social benefit and economic benefit. With the development of our social economy, the demand for power in increasing. Power load can also be affected by date, weather, climate, economy and politics, all of which increase the difficulty of precise load forecasting.This thesis first introduces the background of power load forecasting and the influence of power forecasting to economy. We analyzed the factors affecting the power load and basis model of power analysis. We introduced several base methods which are used for power load forecasting. We focus on the usage of artificial neural networks (ANN) in power load forecasting and introduced the principle of ANN. We also proposed a new ANN method for load forecasting based on serial incremental ratio. The main contributions of this thesis includes following three aspects:Firstly, we introduced the basic models and basic methods of power load forecasting. This thesis introduced the influence of four basic components, which includes normal part load component, weather-sensitive load component, special event load component and random part load component, by visualizing the experimental data to discover the individual effect of four component to the variation trend of total load. Then we introduced several methods of load forecasting, including traditional analytical prediction method, regression methods, gray prediction and neural networks.Secondly, we introduced the basic principles of neural networks. We introduced two kinds of neural networks, basic feed-forward neural networks and recurrent neural networks for load forecasting. We introduced how to learn network parameters by back-propagation. We introduced the framework we used for load forecasting, including data processing, data normalization, feature quantization, and how to determine the number of hidden units. We proposed our load forecasting method based on serial incremental ratio.Thirdly, we test our framework by experiment conducted on mentioned two kinds of neural networks. For every network, we set two structures by comparing basic load prediction and serial incremental ratio prediction. Since RNNs have the ability of history memory, we only add the previous history load in the network input in methods of feed-forward neural networks. Our experiment adopted the dataset in 2001 World-wild competition within the EUNITE (EUNITE, the European Network of Excellence on Intelligent Technologies for Smart Adaptive Systems) network. This competition adopted the power load data of Slovakian in year 1997 and 1998 provided by Eastern Slovakian Electricity Corporation to predict the power load of January in 1999. We adopt Mean-Absolute-Precision-Error and Max-Error as evaluation metrics. Extensive experiment showed that our method achieve obvious improvement in precision of load forecasting compared with base methods.