Research On Improved Wind Speed Forecasting Based On NARX And Key Technologies Of Wind Power Forecasting

Accurate wind speed and wind power forecasting is the key to ensure the safe and stable connection of wind energy into the power grid system.Accurate wind speed forecasting can avoid the unnecessary loss of wind turbines due to high wind speed,and it is also helpful for wind power forecasting.However,the wind speed data is different from the general time series data.The intermittency of the wind has the characteristics of large fluctuation and weak continuity,which makes it exist a large number of peak points within the range of higher wind speed.These peaks make it difficult to predict wind speed accurately.The existing research has not proposed an improved method for the feature of wind speed.Accurate wind power forecast can provide reliable reference for power supply planning and energy dispatching in power market.However,the relationship between wind speed and wind power is severely distorted by the large number of anomalies in the wind power data set acquired by the Supervisory Control And Data Acquisition(SCADA)system,especially the sparse anomalies distributed around the curve.Removing abnormal points will destroy the continuity of the time series data.Although the missing values can be filled by interpolation method,it will also lead to distortion of wind power data set when the elimination rate is too high.At the present stage,on the premise of classifying and discussing abnormal points,comprehensive analysis has not been conducted on the above problems existing in Wind Turbine Power Curve(WTPC)model.In view of the above problems in wind speed and wind power forecasting,this paper puts forward C-NARX network and Quartiles-DBN model respectively.Its innovations and main works are as follows:To solve the problem of low prediction accuracy at high peak wind speed,a NARX wind speed forecasting network based on Copula,C-NARX,was proposed.The construction process of C-NARX network is as follows: First,the probability distribution function modeling of target wind speed is completed through analysis,and appropriate meteorological data is selected as additional input of NARX network through nonlinear coefficients.Then,according to the nonlinear correlation and tail dependency,the appropriate connection objects are selected,and the probability distribution function modeling of the connection objects is completed.Then,the appropriate Archimedes Copula connection function is selected on the basis of the above,and the construction of the joint probability distribution function between the target wind speed and the connected object is completed.Then,the value of the connection object under the p-quantile of the joint probability distribution function is regarded as the threshold.The threshold value is used as a criterion to predict whether the target will reach the peak value in the high wind speed range,so as to complete the construction of the new column,copula＿input.Finally,the copula＿input,together with the target wind speed and additional meteorological inputs,is treated as an input Tensor to complete the training of the network.The experimental results show that C-NARX network has a good prediction effect,which is better than similar NAR/NARX based networks.In view of a large number of outliers in wind power data set and the problem of data set distortion after removing outliers,this paper proposes a Deep Belief Network(DBN)wind power curve model Quartiles-DBN based on Quartiles algorithm.The model is used to realize wind power forecasting,and its construction process is as follows: the Quartiles outlier detection algorithm firstly classifies the wind speed and wind power(v-p)points,and removes the deviation points except the sparse outlier points to complete the preliminary cleaning of the data set.Then,the Quartiles algorithm detects the sparse outliers.Sparse outlier detection algorithm consists of two parts: vertical detection method and horizontal detection method.It is divided from two data dimensions of wind speed and wind power respectively,and the normal value range of points within each partition interval is calculated.It completes all the preprocessing work of the whole data set by the way of culling while sliding.Then,WTPC is constructed by DBN deep neural network.Since DBN is insensitive to the discontinuity of time series data,it solves the problem of incomplete data set after removing a large number of outliers and the problem of distortion after interpolation.Finally,the wind speed forecasting value is brought into the WTPC to realize the wind power forecasting under the same time scale.The experimental results show that the Quartiles-DBN model has good wind power forecasting accuracy and can complete the task of wind power forecasting.Finally,a wind farm management system SMART＿W based on the above network and model is developed to meet the requirements of wind farm production process.The test results show that the network and model proposed in this paper have high prediction accuracy.