Day-ahead wind power combination forecasting based on corrected numerical weather prediction and entropy method

To satisfy the grid operation scheduling requirements for wind power forecasting model accuracy, the measured wind speed near the height of the wind turbine hub is added to the wind power combined forecasting model. First, the relationship between the numerical weather prediction wind speed and the measured wind speed at different heights are analysed, and the correlation between each wind speed and the wind power is compared. Second, the random forest algorithm combined with the cumulative contribution rate is used to select several meteorological types of numerical weather prediction data as the input of the long short-term memory network to predict wind speed. Third, while inputting the meteorological data provided by numerical weather prediction, which is highly related to wind power, the wind power prediction network also uses the predicted wind speed of the upper network as input to predict wind power. Finally, the entropy method is used to dynamically determine the combined weights of each forecasting model and improve the adaptability of the model. Research and analysis using measured data from two wind farms located in northeast China have veriﬁed the effectiveness of the method.


INTRODUCTION
The large scale of wind power access to a power grid results in a series of uncertainties to the rational dispatching and safe and stable operation of power systems involving wind power. To solve the grid scheduling problem caused by the intermittent and volatile nature of wind power, the short-term prediction of wind power can be made more accurate; more accurate prediction effectively reduces the rotating reserve capacity of the grid and not only reduces the operating cost of the power system but also provides a reliable basis for power system scheduling and stable operation [1,2].
Based on the different requirements of the power system for wind power integration, the prediction scale can be divided into four types. Real-time prediction, that is, ultra-short-term multi-step prediction, mainly uses the measured historical wind power and wind speed data modelling, and the rolling prediction is 15 min in the next 4 h [3]. Here the wind power is convenient for real-time control of the operating state of the wind turbine; In day-ahead prediction, a variety of numerical weather prediction information is collected for modelling, and the wind power for 24 h each day is predicted to meet reasonable dispatching of the power grid and the bidding requirements of the wind power [4,5]. The third type, medium-term prediction is based on the weekly or monthly forecast period and is mainly used for the regular maintenance and commissioning of wind farms. Long-term prediction with an annual period is mainly used for the feasibility analysis of wind farm construction.
At present, the methods of short-term wind power prediction can be divided into physical methods, statistical information methods, and artificial intelligence methods [6]. The physical method requires a large amount of data such as topographic information and meteorological information. Although it has high prediction accuracy, its computational cost is high [7].The statistical information method needs to model the complex relationship between historical wind speed and power, but has limited precision [8]. The artificial intelligence method is more and more applied in the field of data prediction because of its selfadaptability and flexibility [9][10][11].
Numerical weather prediction (NWP) data is a critical factor in short-term wind power forecasting [12]. The wind power prediction model is considerably influenced by the volatility and spatial-temporal correlation of wind, and this correlation is strongly dependent on factors such as wind direction, wake, terrain roughness, and ground height [13]. However, NWP theoretically fails to incorporate incentives related to wind volatility and spatiotemporal correlation. Sideratos [14] proposed that the error between NWP wind speed and measured wind speed follows a certain law.
Considering that common physical methods can hardly eliminate the limitations of system errors, as well as the complexity of calculations required for uncertain initial conditions and estimated meteorological parameters. Many researchers began to focus on data-driven methods and modified the NWP based on historical wind data. Yan [15] established an NWP correction model with multiple inputs and multiple outputs by analysing the deviation patterns between the NWP data and the measured data of the wind tower. However, it only reduces the dimension of NWP and does not improve the accuracy of NWP wind speed prediction. Wang [16] proposed an ultra-short-term wind power prediction method based on OS-ELM, realizing the real-time correction of NWP wind speed and rapid prediction of wind power. There is no in-depth analysis of the correlation between NWP wind speed and SCADA wind speed. Buhan [17] clustered the NWP grid information, and for each cluster, an artificial neural network or support vector machine (SVM) model is constructed to learn the relationship between the wind patterns of NWP data and wind speed data measured by wind tower. It does not systematically analyse the relationship between the data at different heights. Du [18] proposed a method to predict wind power, which uses both NWP data and meteorological observation data. However, it did not systematically analyse the relationship between the information provided by the NWP and the measured meteorological information, and did not make a reasonable explanation at the physical level.
Fan [19] proposed that meteorological data of different heights can better characterize the atmospheric characteristics around the wind farm, and using it as an input to the prediction model can reduce the prediction error. However, the method used is a single-layer artificial neural network, which cannot fully learn the relationship between multi-source information. At present, the accuracy of a single prediction model has been difficult to improve, and the study of combined models is a trend to make up for this defect. The weighting of the combined model is mainly determined by the weighted value method. The prediction results cannot fully consider the main information of each model, and there is optimization space for prediction accuracy [20]. Chen [21] proposed using the crossentropy theory to determine the weight of the combined forecasting model, and the prediction accuracy is improved, but the models' adaptability is not high and lacks online modelling ability. Deep learning methods can learn abstract features in a large amount of data through multi-layer nonlinear modelling, and can better model the nonlinear relationship between multi-source information. Recently, the long short-term memory network (LSTM) have been applied to short-term prediction.
Based on the analysis of the correlation between measured wind speed data and corresponding NWP data at different heights of wind farms, a short-term wind power combined forecasting model based on the LSTM-entropy method (EM) is proposed. First, the prediction model inputs the meteorological information provided by some types of NWP into the LSTM network to predict the measured wind speed and implement NWP correction. Then, based on the forecast wind speed at different heights output from the upper network, the LSTM-based wind power forecasting model is established online and dynamically determines the prediction results via the entropy method. Finally, the weights are superimposed to get the final predicted value. The method can be modelled and predicted online, and the calculation speed is fast and has certain practical engineering value. The actual example shows that the LSTM-EM combination model has higher prediction accuracy than the NWP direct prediction and average weight method, and the prediction results of LSTM are better than those of traditional SVM and back propagation (BP) neural network. At the same time, we also compared the LSTM with the gated recurrent unit (GRU). When processing large amounts of data, LSTM still has obvious advantages.
The rest of this paper is organized as follows. Section 2 introduces the principles of entropy method and LSTM. Section 3 introduces the reason and method of NWP wind speed correction (use NWP data to predict measured wind speed). Section 4 establishes a day-ahead combination forecasting model for wind power based on LSTM and entropy method. Section 5 verifies the effectiveness of the proposed model through two calculation examples, and the conclusion is given in Section 6.

Principle of entropy method
In information theory, entropy can measure the degree of disorder of the system and obtain the contribution value of useful information as a whole, determining the weight of the total information and providing the basis for decision-making of the comprehensive decision-making evaluation system. The larger the amount of information provided by the indicator, the smaller the entropy of the indicator and the greater the weight and its role in the comprehensive evaluation system. The entropy method can exclude the influence of subjective factors. According to effective information, the degree of difference between the indicators can be objectively evaluated and intuitively reflected.
An evaluation system with n evaluation indicators of m objects is set up, and the evaluation matrix of its composition is X = (x i j ) m× n , i = 1, 2, … , m; j = 1, 2, … , n, the standardization method is as follows: In the formula, the standardized data of P i j after standardization can eliminate the difference between the indicators. The entropy of each evaluation indicator is defined as: Assume that when P i j = 0, let P i j ln P i j = 0. On this basis, the entropy weight of each evaluation index j is defined as: Among them, the sum of the entropy weights j is 1, which embodies the information amount of each evaluation index. The larger the value, the greater the effect of the index on the comprehensive decision.

Principle of LSTM
LSTM neural network is a deep learning method that can make full use of training samples and can deeply mine the complex information in sample data to complete the approximation of high-dimensional functions [22]. As a typical time series, wind power is not only nonlinear but also has dynamic characteristics; that is, the output of the system is related not only to the input at the current time but also to past inputs. LSTM has a unique memory and forgetting mode that flexibly adapts to the timing characteristics of network learning tasks. At the same time, the LSTM network solves the problem of gradient disappearance and gradient explosion in the training process of recurrent neural networks, and can make full use of historical information to model the time dependence of signals [23]. LSTM solves the long-term dependence of traditional neural networks in time series problems. By introducing memory cells to model the long-term dependence in the original signal, it shows the effect of long-term memory. With deep learning ability, it is suitable for prediction and can deal with important events with long delays and long intervals in the time series. The LSTM includes an input layer, an output layer, and a plurality of hidden layers. The hidden layer is composed of a plurality of storage units, and its basic structure is as shown in Figure 1. Each memory unit is composed of four parts, including a memory cell, an input gate, a forget gate, and an output gate. The number of hidden layers and the number of neurons they contain determine the effectiveness of the network training. The more hidden layers and more hidden nodes in each layer, the more complex the model, and overfitting may occur. Conversely, if the hidden layer nodes are too small, the network performance is worse.
In the figure, one LSTM unit contains a cell, and c t indicates the state of the hidden layer at time t . This tuple is regarded as the memory unit of the LSTM. x t denotes the input of the memory unit at time t , o t denotes the output of the memory unit at time t , and f t denotes the output of the forgetting gate at The specific workflow of the LSTM unit is as follows: at time t , the LSTM unit receives the input of two kinds of external information of the current state x t and the hidden state h t −1 of the previous time LSTM through three control gates. In addition, each control gate also receives an input of information inside the memory unit, that is, the state of the memory unit c t −1 . After receiving the input information, each control gate operates on the input from different sources through the activation function to determine whether it is activated. After the input of the input gate is transformed by the activation function, the state of the memory cell processed by the forgetting gate is superimposed to form a new memory cell state c t . Finally, the memory cell state c t forms the output h t of the LSTM unit by the operation of the activation function and the dynamic control of the output gate. The specific calculation formula between the variables is as follows: Where: W xi , W x f , W xc , W xo is the weight matrix input to the hidden layer; W hi , W h f , W hc , W ho is the weight matrix outputting the hidden layer; W ci , W c f , W co is the diagonal matrix connecting the neuron activation function output vector c t and the gate function; b i , b f , b c , b o is the offset vector; g is the sigmoid activation function.

Comparison of NWP wind speed and measured wind speed
The data provided by NWP is an important input for shortterm wind power forecasting models. NWP data contains a variety of meteorological information, such as wind speed, wind direction, temperature, and air pressure. Wind speed and wind direction are important factors affecting wind power generation. NWP generally provides wind speed and wind direction information at different heights. However, the height of the wind speed and direction provided by the NWP may be different from the height of the wind turbine hub. This leads to errors between the wind speed provided by the NWP and the wind speed measured by the wind farm. Even if the height of the wind speed provided by the NWP is consistent with the height of the measured wind speed, the existence of prediction errors cannot be avoided. Therefore, the NWP can be corrected to improve the accuracy of the NWP data and achieve the purpose of improving the accuracy of wind power forecasting.
To analyse the correlation between the measured meteorological data and the NWP data, the wind power related data of a large-scale wind farm in Northeast China is taken as an example, with the data time range from January 1 to August 8, 2017 and a time resolution of 15 min. The data of the wind tower includes wind speeds of 10, 50, 65, and 80 m and wind direction data. The wind farm has a total of 267 MY1.5SE-1.5 MW units with a total installed capacity of 400.5 MW and a hub height of 80 m. Historical wind power data is preprocessed: data less than 0 is replaced by 0, and data larger than the total installed capacity is replaced by 400.5 MW. NWP data includes wind speed and wind direction data at heights of 10, 30, 100, and 120 m, as well as a large amount of meteorological data such as temperature and pressure, a total of 24 items. NWP is provided every 12 h and contains forecast data for the next 72 h. We connect the data of the previous 12 h in each NWP forecast result to obtain continuous NWP data. Figure 2 shows the fluctuation curve of the measured wind speed and NWP wind speed of the wind farm. As shown in Figure 2, the fluctuation trend of the wind speed curve provided by NWP at different heights is consistent. NWP gets the meteorological forecast results by solving the hydrodynamics and thermodynamics equations, and then gets the meteorological forecast data at various heights by scaling down. Therefore, the information contained in the wind speed and wind direction of NWP at different heights is similar, and adding NWP FIGURE 2 Comparison of NWP wind speed and measured wind speed sequence data at different heights to the prediction model will increase the redundancy of the model and the difficulty of training. It is sufficient to analyse only the NWP data near the height of the hub. It can be seen from Figure 2 that the wind speeds measured at heights of 80 and 65 m are similar, but the wind speeds measured at heights of 50 and 10 m are quite different from the previous two. This is because when the altitude is low, the wind speed is more affected by the soil roughness. Although the wind speed at low altitude contains some useful information, it contains more interference information. If the wind speed at a low altitude with weak correlation with wind power is used as the input of the prediction model, the accuracy of the prediction will be reduced.

Correlation verification of wind speed and wind power
To verify the reasonableness of the qualitative analysis, a quantitative analysis is performed below. From the perspective of linear correlation, the correlation between wind power and various wind speeds is analysed using Pearson correlation. The correlation coefficient is expressed by the following formula: In the formula, is the correlation coefficient, E is the expected value, V is the variance, X P is the wind power, X V is the wind speed.
The statistical results show that the correlation coefficients between wind energy and the measured wind speeds of 80 and 65 m are 0.739 and 0.728, respectively, and the correlation coefficients with 10 and 50 m are 0.627 and 0.639. The correlation coefficients between wind energy and the NWP wind speed of 80 m is 0.538, and the correlation coefficients with 10, 30, and 120 m are 0.468, 0.496, and 0.523, respectively. This statistical result confirms the conclusion of our above-mentioned physical analysis. In the next work, we mainly use the 100 m NWP wind speed data to predict the 80 and 65 m wind speeds, which can also be regarded as a correction to the NWP wind speed data.
In addition to wind speed and wind direction data at different altitudes, NWP also provides various meteorological information. If all the NWP information is used as the input of the correction model, it is difficult to avoid the negative effects of data redundancy. Therefore, we use the Random Forest (RF) algorithm to calculate the importance of various NWP data to the measured wind speed, and then select the main meteorological type through the cumulative contribution rate. In order to consider the importance of various NWP data to the measured wind speed, RF randomly initializes each NWP data type item by item. If a certain type of NWP data is more important for the measured wind speed, after random initialization of this data, the model error will increase and the mean decrease in accuracy (MDA) will decrease. In order to minimize the model generalization error, the number of decision trees and the number of candidate data types need to be set in advance. Then sort the MDA in descending order and calculate the cumulative contribution rate of the NWP data. The calculation formula of the cumulative contribution P is as follows: In the formula, v represents a certain kind of NWP data, N is the total number of meteorological data types provided by NWP (wind speed and wind direction data only take the data at 100 m altitude). M(v) is the MDA of NWP data v. p represents the number of accumulated NWP data types.
Take the data of the previous 180 days of the wind farm as an example. The number of decision trees is 1000, and the number of candidate NWP data types for each node is four. When the cumulative contribution rate is greater than 80%, the corresponding p kinds of NWP data are used as the input of the correction model. As the dimensions of all meteorological data are different, normalization is required. The wind direction data takes its sine and cosine values as normalized values, and other types of NWP data are divided by the historical maximum, respectively. The input vectors of the final selected correction model include wind speed, wind direction, air pressure, temperature, humidity, and momentum flux.
As a deep learning algorithm, LSTM has excellent information learning capabilities, and can well model the nonlinear relationship between multiple inputs and a single output. We use LSTM to simulate the relationship between the NWP data and the measured wind speed to predict the wind speed of 80 and 65 m. The network structure diagram of the NWP wind speed correction model is shown in Figure 3. The input layer is the normalized meteorological data of the NWP at a certain time point, and the output layer is the measured wind speed of the wind tower normalized at the same time point.

WIND POWER FORECASTING MODEL BASED ON LSTM AND ENTROPY METHOD
In theory, the numerical weather prediction information can make up for the shortcomings caused by the historical wind speed or power data directly predicting the wind power, but the existing numerical weather forecast report data and the wind farm measured wind energy data have certain differences. In the previous chapter, we proposed a correction method for NWP wind speed, using NWP data and measured wind speed, to predict the wind speed near the height of the wind turbine hub. In addition, the terrain of the wind farm is uneven and it is impossible to ensure that the wind turbines are installed at the same height. The measured wind speed data at the height of the hub of the wind turbine contains more information, but the measured wind speed data at other heights also contains some useful information. And the wind speed measured at a lower altitude has a weak correlation with wind power and cannot be used for wind power prediction. Therefore, we only add the measured wind speed near the height of the wind turbine hub to the wind power prediction model.
In view of these features, a proposed method for predicting wind power based on the LSTM and entropy method is proposed. First, by analysing the relationship between the wind energy data of the wind farm, select the wind speed height to be used. Then, the LSTM network of the first module is constructed to correct the NWP wind speed to obtain the predicted wind speed at the target height. In the next step, the second module of the LSTM network is constructed. While inputting NWP data that is highly correlated with wind power, the corrected NWP wind speed output by the network of the previous module is also used as input to predict wind power.
The optimal model parameters are selected based on the characteristics of the NWP wind speed correction model and the power prediction model. Finally, the entropy method is used to determine the weight of each models' predicted value, 1. Analyse the relationship between NWP data, measured wind speed and wind energy. Choose the height of the wind speed with higher correlation with wind energy as the target height. The target height H i with higher correlation degree between wind power and the measured wind speed is determined, where i ≤ n, n is the number of wind heights measured by the wind tower. 2. Based on the RF algorithm combined with the cumulative contribution rate, the NWP data that is highly related to the measured wind speed at the target height is selected as input, the wind speed data at the target height H i is used as the output to establish an LSTM-based NWP wind speed correction model, the optimal parameters are selected, and then the corrected NWP wind speed data is obtained. 3. In the second prediction module, while the NWP meteorological information other than wind speed, which is highly related to power, is used as the model input, the corrected NWP wind speed output by the upper prediction module is also used as input, and the corresponding historical power data as the output to establish the wind power forecast model selecting the best parameters of the model and obtaining the power prediction results. 4. The weight of each power prediction result is obtained by the entropy method, and the final wind power output combination prediction result is obtained by weighting. Among them, the object m to be considered when calculating the entropy weight is the 96 predicted values, and the evaluation index n

Experimental data description
The data used in the study is the same data source as in Section 2. A total of 220 days of data from January 1 to August 8, 2017 are considered. The data used in the study is the same data source as in Section 2. A total of 220 days of data from January 1 to August 8, 2017 are considered. The data sampling interval is 15 min, and one day contains 96 groups of data. In the field of wind power forecasting, many information sources will change over time, and dynamic training can make the model adapt to changing data. We keep a fixed size of training set (180 days), when the new day's data is input to the model, the training set will discard the first day's data. We use the first 180 days of 220 days as the initial training set, and the last 40 days of 220 days as the test set. For the actual situation of the proposed short-term wind power prediction model, the data structure of the example can be represented by Figure 5.
In this study, the LSTM-based NWP wind speed correction model and power prediction model are built by Theano as the back end through the Keras deep learning framework of the Python platform. The optimal parameters of each model after multiple simulation training are as follows: the number of layers in the two model networks is 3, the number of nodes in each layer is 7, 17, and 1, respectively, and the number of iterations is 160 and 200 respectively. The Adam optimizer is used to train the model, and the batch volume is 1. The number of iterations is different because the types of sample data of the two models are different and the model needs to continuously update the data to model online.

Evaluation index
In order to ensure an effective and comprehensive evaluation of the accuracy of the NWP wind speed correction, we have selected two different evaluation indicators: mean absolute error (MAE) and root of the mean squared error (RMSE).
In the formula, N is the number of test samples, V t is the wind speed measured at time t, andV t is the predicted wind speed at time t , that is, the corrected NWP wind speed.
The evaluation indicators for wind power prediction effects include: MAE and three evaluation indicators (RMSE, accuracy r1, and pass rate r2) specified in the Interim Measures for Short-Term Wind Power Forecast Management in the National Energy Administration document. The maximum error of the daily prediction curve provided by the wind farm power prediction system shall not exceed 25%, and the root mean square error of the all-day prediction result shall be less than 20%.

Case analysis
Based on the results of the wind energy data analysis of the wind farm in Section 2, it is determined that the target heights of the wind towers of this example are 65 and 80 m. The NWP wind speed correction models based on target heights of 80 and 65 m were established, named V65 and V80, respectively. Figure 6 shows the corrected NWP wind speed fluctuation curve on the 20th day of the training set. It can be seen from the figure that the corrected NWP wind speed is closer to the measured wind speed. Table 1 shows the wind speed prediction accuracy of the revised NWP wind speed at the target height. After the NWP wind speed is corrected, the prediction accuracy has been significantly improved.  Table A1 shows the NWP wind speed correction results after replacing the basic network LSTM with other networks. It can be seen from the table that the NWP wind speed correction model based on LSTN has the highest prediction accuracy.  Based on the NWP wind speed correction model, the corresponding wind power forecast models were established, named P80 and P65, respectively. The entropy weights of the combined model dynamics are determined based on the final prediction results of the two power prediction models. Table A2 shows the average results of entropy weights obtained from the 40-day test days of each prediction model. It can be seen from Table A2 that the prediction result of P80 for power prediction after correcting the NWP wind speed by the wind speed of the wind tower 80 m is more accurate. It shows that the data close to the height of the hub is more suitable as the input data for the power prediction model.
In order to verify the effectiveness of NWP wind speed correction, the conventional method of directly using the NWP data with higher power correlation as the model input was compared with P65 and P80. The error statistics of the prediction results of each method are shown in Table 2. It can be seen from Table 2 that using the corrected NWP data as the model input can significantly improve the accuracy of wind power prediction. Compared with the conventional model, the accuracy and pass rate of P65 are increased by 1.71% and 3.92%, respectively.  The accuracy and pass rate of P80 have increased by 2.18% and 5.44%, respectively. The prediction effect of P80 is better than that of P65, which is consistent with the conclusion in the previous paragraph.
It should be noted that the forecast accuracy of wind farms at high output levels is generally lower than that at low output levels. The reason for this result is that the wind speed fluctuations when the wind farm output is high are too complicated, which makes LSTM unable to learn how to correct the complex fluctuations of the NWP wind speed. A further study can be conducted on how to improve the robustness of the LSTM modified model under complex weather conditions.
In order to verify the validity of the proposed dynamic combination prediction model based on entropy weight, the equal-weight combination of P65 and P80 was compared with the conventional method and the proposed method. Figures 7  and 8, respectively show the prediction results on the first and last day of the test set, and Table 3 gives the corresponding statistical errors. It can be seen from Table 3 that the accuracy of the two types of combined prediction models is higher than that of the NWP direct prediction model. Compared with the equal weight model, the accuracy and pass rate of the proposed model are increased by 0.11% and 0.43%, respectively. Although the degree of precision improvement is not high, the advancement of the entropy method has been verified to some extent.   average errors of all the test days of each model are listed in Table 4. It can be seen from Figures 9 and 10 that the power prediction curve of the LSTM-EM and the GRU-EM are closer to the actual curve of the wind farm for the same test sample. From the average error of all test days given in Table 4, the LSTM model can be seen that compared with the BP neural network and the SVM model, the accuracy of the proposed prediction method is increased by 2.58% and 1.92%, respectively, and the pass rate is increased by 4.36% and 3.80%, respectively. The MAE of LSTM-EM is similar to that of GRU-EM, but the RMSE of LSTM-EM is 0.6% lower than that of GRU-EM.
Taking another wind farm with an installed capacity of 200 MW in Northeast China as an example to further discuss the effectiveness of the proposed method. The time frame of the data used is from January 1st to December 31st, 2018. Use the data of 120 days before March 1 as the training set and the data of the next 30 days as the test set to predict the power in March. In June, September, and December, the power forecast is also carried out according to the above method. The conventional method that NWP is directly used as the input of the model is compared with the proposed method, and GRU and LSTM are respectively used as the basic model to further compare the performance of each model. The prediction results of the 20th day of each test set are shown in Figures A1-A4. Table 5 shows the statistical results of the prediction errors of each model. It can be seen from the prediction results that the LSTM-EM is better than the LSTM. RMSE was reduced by 1.5% in the best case (June) and 0.65% in the worst case (March). MAE was reduced by 1.37% in the best case (June) and 0.17% in the worst case (September). The prediction accuracy of the LSTM-EM is higher than that of the GRU-EM, except for December. The GRU is a variant of the LSTM, and its calculation speed is faster than the LSTM, but the prediction accuracy in the case is generally lower than the LSTM. In engineering practice, if there is a need to shorten the running time of the model, GRU can replace LSTM as the basic model. Of course, this may slightly reduce the prediction accuracy of the model, but with the further development of GRU, it may be as accurate as LSTM.
The prediction results show that the root mean square error of the proposed method is lower than the national industry standard for wind power forecasting and meets the practical application of the project. The reason why the accuracy of the prediction method is improved in this study is that the deep learning method can deeply mine the effective information in the multiheight layer wind measurement data, thereby improving the prediction effect. However, when the traditional algorithm adds too many variables that are not closely related, it is easy to cause "confusion" and bias in model training.

CONCLUSION AND OUTLOOK
To improve the prediction accuracy of short-term wind power, based on the LSTM theory, a short-term wind power dynamic combination forecasting method combining modified NWP and entropy method is proposed, and the method is verified using various examples. The following conclusions are drawn: 1. The wind speed provided by NWP may be quite different from the wind speed measured by the wind farm. Using the NWP wind speed as the model input directly amplifies the prediction error of the NWP, resulting in a decrease in the accuracy of short-term wind power prediction. Therefore, it is necessary to correct the NWP wind speed. The NWP wind speed correction method proposed in this paper can effectively improve the prediction accuracy of the measured wind speed at the target height. It is proved that the correction of NWP wind speed can effectively improve the prediction accuracy of short-term wind power. The prediction accuracy of the proposed model method is better when the wind farm has low output. Improving the robustness of the LSTM modified model under complex weather conditions will be the next phase of research. 2. The wind power forecasting method based on LSTM-EM can better model the nonlinear relationship among multisource information through feature learning from a large amount of data. The excellent feature learning ability and modelling ability of LSTM is an important reason for the improvement of wind power prediction accuracy. GRU has a simpler structure than LSTM and runs faster. However, LSTM is more stable when processing large amounts of information. In this paper, LSTM performs better than GRU in predicting ability, but GRU is also a hot model for future research.
3. The proposed indicators of the dynamic combination forecasting model prediction results are better than the prediction results of each comparison model, verifying the advanced nature of the entropy weight method and the practicability of LSTM in the field of wind power prediction. The advantage of the combined model is that it can fully learn the effective information of each height data. Dynamic training enables the model to adapt to changing data. This provides a reference for the in-depth study of short-term wind power forecasting in the future.