Short-term Wind Power Prediction Based on Particle Filter and Radial Basis Function Neural Network

Abstract

In order to improve the accuracy of wind power prediction, this paper proposed a new short-term wind power prediction combined method based on particle filter and radial basis function neural network. First, the historical wind speed data are processed with particle filter, and the processed wind speed data combined the historical wind direction data and temperature data are using as the input data of the forecast model. Then, the PF-RBF neural network of wind power output forecasting model is established according to the new input data. The experimental results show that the proposed forecasting model has a good accuracy for wind power prediction. Introduction Nowadays, with the gradual consumption of coal, oil, and other nonrenewable fossil energy, the development of renewable energy has been widespread concerned in the world. As an important part of renewable energy, wind power is a kind of wide distribution, large reserves, and clean energy, and its development and utilization has been paid the world's attention [1-2]. Meanwhile, the wind power prediction has become a research hotspot. When large-scale wind power is connected to the grid, the instability of the wind power will seriously damage to the stable operation of power system and the power quality [3]. Thus, improving the accurate prediction of wind speed or wind power has an important practical significance on the stable operation of the power system and on reducing the operation cost of power system. During the wind power prediction development process, many methods were produced , such as Auto-Regressive and Moving Average Model [4], Kalman Filtering Method [5], and the method of Artificial Neural Network [6], Support Vector Machine [7] and so on, and these methods had good prediction effect on wind power prediction. However, in recent years some scholars have put forward some combination methods of wind power prediction, which can greatly improve the prediction accuracy in a certain extent. Wang et al. [8] proposing one combined method based on the improved empirical mode decomposition algorithm and neural network, comparing to the single forecast model, and getting a very good prediction effect. Chen et al. [9] made the information fusion technology apply to the wind power prediction, and proposed a wind power forecasting combined method by using cross entropy theory, effectively improving the prediction accuracy. However, we found that the wind speed is often the main reason affecting the power output. The wind speed data getting from the wind farm are often poor regularity, which are usually with bad data. To solve this problem, the paper used particle filter for data treatment, proposing a new combined method based on particle filter and radial basis function neural network. The historical wind speed data are first smoothly processed by the particle filter, and the singular points from the wind turbine operation can be removed. Then, according to the processed wind data, historical wind direction and temperature, we established the wind power prediction model, which is based on the radial basis function neural network. Finally, the prediction results are compared with the actual wind power through the MATLAB platform for experimental simulation. International Conference on Applied Science and Engineering Innovation (ASEI 2015) © 2015. The authors Published by Atlantis Press 188 Particle Filter Algorithm and Realization in Wind Speed The idea of Particle Filter (PF) [10] is a Monte Carlo simulation method, based on principle of Bayesian estimation using a series particle to represent a probability, and it can be commendably used for treatment of nonlinear systems [11]. Historical wind speed data can be described as a series of volatility discrete point set, and through the state space model of the system, particle filtering can make full use of the state values at k-1combined with the present observations to process values at time k. Particle filter is established on the basis of Bayesian estimation, which is mainly to construct the posterior probability density function of the unknown system states by prior knowledge and the actual observation data. The state space model can be written as Process equation: 1 1 ( , ) k k k x f x u − − = (1) Observation equation: ( , ) k k k y h x v = (2) where f()and h() denote the known state function and observation function respectively; xk is a state of the system at time k, uk and vk represents random noise vectors of given distributions; yk represents observation value at time k, uk and vk are independent of each other and independent of the past and current state. The particle filter working principle can be described as the formula (1) and (2). It updates a priori knowledge through the priori probability density and new coming observation value, and obtains the desired posterior probability density. Then, according to Bayesian estimation principle, the posterior probability density can be got by both prediction and update steps. Thus, formula (1) and (2) can be written as