Wind power prediction method and system based on improved neural network quantile regression

Abstract

The wind power prediction method and system based on improved neural network quantile regression disclosed in the present invention include: obtaining weather forecast data for multiple wind farms; and obtaining wind power generation prediction results for each wind farm based on the weather forecast data and a trained wind power generation probability prediction model. The wind power generation probability prediction model uses the weather forecast data for each wind farm as input and the wind power generation prediction results as output. It is constructed using a neural network-expressed quantile regression algorithm, and the activation function of the model output layer uses a ramp function. A smoothing function is introduced into the ramp function and the test function of the quantile regression algorithm. This improves the accuracy of wind power prediction, and the constructed wind power generation probability prediction model has good generalization capabilities.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation technology, and in particular to a wind power prediction method and system based on improved neural network quantile regression. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Different from the power generation forecast of a single wind farm, wind power generation forecast aims to estimate the total power generation of multiple wind farms in a region. It is a supplement to the current wind power forecast model and is of great significance in alleviating the negative impact of wind power grid connection on the safety and stability of the power system.

[0004] At present, the probabilistic prediction model is mainly used to predict the wind power generation capacity of multiple wind farms. However, when predicting the wind power generation capacity, it only relies on the data of a single wind farm without utilizing the related information between the wind farms, resulting in inaccurate prediction results. In addition, the existing model has the problem of non-differentiability at the origin when predicting the wind power capacity, and the generalization ability of the model is limited. Summary of the Invention

[0005] In order to solve the above problems, the present invention proposes a wind power prediction method and system based on an improved neural network quantile regression. A ramp function is introduced in the output layer of the wind power generation probability prediction model to avoid the intersection of quantile curves. The introduction of a smoothing function overcomes the problem that the network's loss function and the ramp function of the network output layer are not differentiable at the origin, thereby improving the generalization ability of the model.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] First, a wind power forecasting method based on an improved neural network quantile regression is proposed, including:

[0008] Obtain weather forecast data for multiple wind farms;

[0009] Based on the weather forecast data and the trained wind power generation probability prediction model, the wind power generation prediction results of each wind farm are obtained. Among them, the wind power generation probability prediction model takes the weather forecast data of each wind farm as input and the wind power generation prediction results as output. It is constructed by using a neural network expression quantile regression algorithm, and the activation function of the model output layer adopts a ramp function, and a smoothing function is introduced in the ramp function and the test function of the quantile regression algorithm.

[0010] Secondly, a wind power prediction system based on improved neural network quantile regression is proposed, including:

[0011] A data acquisition module, used to obtain weather forecast data for multiple wind farms;

[0012] The wind power generation power prediction module is used to obtain the wind power generation power prediction results of each wind farm based on weather forecast data and a trained wind power generation power probability prediction model. The wind power generation power probability prediction model takes the weather forecast data of each wind farm as input and the wind power generation power prediction results as output. It is constructed using a neural network expression quantile regression algorithm, and the activation function of the model output layer uses a ramp function, and a smoothing function is introduced in the ramp function and the test function of the quantile regression algorithm.

[0013] In a third aspect, an electronic device is proposed, comprising a memory and a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are run by the processor, the steps of the wind power prediction method based on the improved neural network quantile regression are completed.

[0014] In a fourth aspect, a computer-readable storage medium is proposed for storing computer instructions. When the computer instructions are executed by a processor, the steps of the wind power prediction method based on improved neural network quantile regression are completed.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] 1. When the present invention predicts the probability of wind power generation through the wind power generation probability prediction model, in addition to utilizing the wind farm data itself, it also utilizes the correlation between multiple wind farm data, thereby improving the accuracy of wind power generation probability prediction.

[0017] 2. The present invention processes the input layer of the neural network through local connection, which reduces the computational burden of the network, reduces redundant learning, and improves the learning efficiency of the model;

[0018] 3. The present invention introduces a ramp function in the model output layer to avoid the occurrence of the quantile curve intersection phenomenon; the introduction of the smoothing function overcomes the deficiency that the model loss function and the ramp function of the neural network output layer are not differentiable at the origin, thereby improving the generalization ability of the model.

[0019] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings in the specification, which constitute a part of this application, are used to provide further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute improper limitations on this application.

[0021] Figure 1 An improved multi-channel neural network structure diagram according to an embodiment of the present invention;

[0022] Figure 2 This is a test function diagram in an embodiment of the present invention;

[0023] Figure 3 This is a ramp function diagram of the output layer activation function in an embodiment of the present invention;

[0024] Figure 4 A modified test function in an embodiment of the present invention;

[0025] Figure 5 A modified ramp function according to an embodiment of the present invention;

[0026] Figure 6 This is a graph showing a change trend of the reliability index of the wind power generation power at 80% and 90% confidence intervals versus the prediction duration in an embodiment of the present invention;

[0027] Figure 7 This is a graph showing the change trend of the sensitivity index of the 80% and 90% confidence intervals of wind power generation power versus the prediction duration in an embodiment of the present invention;

[0028] Figure 8 This is a graph showing the 72-hour forward-looking forecast results for November 12, 2016, using the improved multi-channel neural network quantile regression forecasting model in an embodiment of the present invention.

[0029] Figure 9 This is a 72-hour forecast result diagram for November 12, 2016, using the multi-channel neural network quantile regression forecasting model in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0031] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0032] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0033] Example 1

[0034] In this embodiment, a wind power prediction method based on an improved neural network quantile regression is disclosed, comprising:

[0035] S1: Obtain weather forecast data for multiple wind farms.

[0036] S2: Based on the weather forecast data and the trained wind power generation probability prediction model, the wind power generation power prediction results of each wind farm are obtained. The wind power generation probability prediction model takes the weather forecast data of each wind farm as input and the wind power generation power prediction results as output. It is constructed by using a neural network to express the quantile regression algorithm, and the activation function of the model output layer adopts a ramp function, and a smoothing function is introduced in the test function of the ramp function and the quantile regression algorithm.

[0037] The wind power generation probability prediction model takes the weather forecast data of each wind farm as input and the wind power generation prediction result as output. It is constructed by using the neural network expression quantile regression algorithm. The specific wind power generation probability prediction model is shown in formula (1), and the loss function of the model is shown in formula (2):

[0038]

[0039] minE{ρ τ (y,Q τ )}(2)

[0040] Among them, h j is the hidden layer data of BP neural network, w j is the weight from the jth neuron in the hidden layer to the output neuron, b2 is the output layer bias, φ(·) is the activation function of the output layer, ρ τ is the test function in the quantile regression algorithm, Q τ is the quantile at the required τ level, and y is the predicted value of wind power generation of the wind farm.

[0041] The neural network disclosed in this embodiment adopts an improved multi-channel neural network, and its specific structure is as follows: Figure 1 As shown, it includes multiple input layers, each of which is connected to multiple hidden layers connected in sequence. The input of each input layer is the weather forecast data of a wind farm. The number of input layers is the same as the number of wind farms. The weather forecast data of each wind farm is input into the corresponding input layer, and then input into the hidden layer through the input layer. Feature extraction is performed through the hidden layers connected in sequence, and the extracted features are output through the output layer.

[0042] The neural network's input layer uses a locally connected approach, with input data categorized by wind farm. Neurons with different superscripts represent inputs from different wind farms. During the training process from the input layer to the hidden layer, there is no cross-learning between data from different wind farms. This process can be considered feature extraction of weather forecast data within each wind farm. The subsequent hidden layer is fully connected. This can be considered as performing correlation learning between wind farms after the first hidden layer extracts key internal features within the wind farms, extracting correlated features from the data between wind farms. This approach reduces redundant computation from the output layer to the hidden layer, alleviating computational burden.

[0043] Specifically, the expression for feature extraction between the neural network input layer and the first hidden layer is shown in (3):

[0044]

[0045] in, is the value of the jth neuron in the first hidden layer of the i-th wind farm, N represents the number of explanatory variables selected for each wind farm, represents the weight of the hidden layer, b j1 Indicates bias, is the activation function, usually the sigmoid function, is the nth input value of the i-th wind farm. The second hidden layer is expressed as formula (4):

[0046]

[0047] By analogy, the expressions of subsequent hidden layers can be obtained.

[0048] The output layer can be expressed as formula (5):

[0049]

[0050] in, is the output layer weight, b jn is the output layer bias, and φ(·) is the activation function of the output layer. τ is the quantile at the required τ level.

[0051] Quantile regression can easily lead to the problem of quantile curve intersection. The quantile should increase with the increase of the probability level τ, and the predicted value should also increase with the increase of the probability level τ. Therefore, the activation function of the model output layer can be replaced by a ramp function, for any 0<τ1<τ<1:

[0052]

[0053] Among them, r(Q τ) is the slope function of the quantile regression algorithm, Q τ is the quantile at the required τ level, Q τ1 is the quantile at the required τ1 level.

[0054] Due to the high uncertainty of wind power forecast, a single quantile cannot represent comprehensive information and cannot provide comprehensive and reliable information for grid operation decision-making. Therefore, multiple quantiles can be used to represent the distribution characteristics of the wind power forecast values ​​of each wind power plant in the wind power cluster, that is, F:

[0055] F={Q τ |0≤τ1<τ2<…<τ r ≤1} (7)

[0056] Gradient descent algorithm is the most widely used optimization algorithm for neural networks. However, it can be seen from formula (6) that the test function of the quantile regression algorithm is not differentiable at the origin, which leads to the model’s loss function not being differentiable at the origin. At the same time, the activation function of the output layer is not differentiable at the origin, which obviously affects the optimization effect of the model. The test function graph and the output layer activation function - ramp function graph are shown as follows: Figure 2 、 Figure 3 shown.

[0057] In order to enhance the generalization ability of the model and make the neural network more compatible with the quantile regression algorithm, this embodiment introduces a smoothing function into the ramp function and the test function of the quantile regression algorithm, and reasonably modifies the test function and the ramp function. The modified test function graph and the modified ramp function are as follows: Figure 4 、 Figure 5 shown.

[0058] The smoothing function introduced at the origin in this embodiment is:

[0059]

[0060] After introducing the smoothing function into the test function, the corrected test function is obtained:

[0061]

[0062] After introducing the smoothing function into the ramp function, the corrected ramp function is obtained:

[0063]

[0064] After introducing the smoothing function into the loss function, the loss function of the wind power generation probability prediction model is obtained as follows:

[0065]

[0066] Among them, r′(Qτ ) is the modified ramp function, ρ τ ′(y,Q τ ) is the corrected test function, h(y,Q τ ) is the introduced smoothing function, E τ ′ is the loss function of the prediction model, N is the number of samples, and τ is the probability level.

[0067] The process of obtaining the trained wind power generation probability prediction model is as follows:

[0068] Obtain weather forecast data and corresponding wind power generation data for multiple wind farms within a preset time period, and construct training and validation sets based on the acquired data;

[0069] The constructed wind power generation power probability prediction model is trained using the training set, and the training termination condition of the wind power generation power prediction model is determined using the validation set. When the training is terminated, the trained wind power generation power probability prediction model is obtained.

[0070] The wind power prediction method based on improved neural network quantile regression disclosed in this embodiment combines the powerful data processing and nonlinear fitting capabilities of multi-channel neural networks with the non-parametric probability fitting capabilities of quantile regression, effectively combines multi-channel neural networks and quantile regression algorithms, and proposes a wind power generation probability prediction model based on multi-channel neural network quantile regression. Then, taking into account the characteristics of wind power prediction of multiple wind farms, a series of improvements were made to the model. During the improvement process, the input layer of the network adopts a local connection method to meet the challenges of big data; the activation function of the network output layer adopts a ramp function to overcome the intersection of multiple quantile curves; and the introduction of a smoothing function makes the model loss function differentiable everywhere, thereby enhancing the generalization ability of the model. The model uses the initial data to directly fit the wind power generation power, fully utilizing the rich information contained in the data and avoiding the generation of cumulative errors from mapping the power generation prediction results of a single wind farm station to the wind power generation prediction results.

[0071] This embodiment uses two probability prediction evaluation criteria to evaluate the proposed wind power generation probability prediction model.

[0072] 1) Reliability

[0073] The confidence interval with a confidence level of 1-2τ is defined as [Q τ ,Q 1-τ ](0<τ<0.5), assuming that the number of training samples is N, the prediction result coverage PICP under the confidence level 1-2τ is:

[0074]

[0075] in, is the characteristic function, for The function value is 1, for The function value is 0. Obviously, the frequency of the sample falling within the confidence interval should be close to the confidence level, so the reliability index of this embodiment is defined as:

[0076] RACE=PICP-(1-2τ) (13)

[0077] The smaller the value of the reliability index RACE, the more reliable the result is, and it is 0 under ideal conditions.

[0078] 2) Sensitivity

[0079] This example evaluates the sensitivity of the model using the confidence interval mean width PINAW, as shown below:

[0080]

[0081] Where N is the number of samples and R is the difference between the maximum and minimum values ​​of the samples. The smaller the PINAW value, the higher the sensitivity of the model.

[0082] To rationally construct the model, this example divides the dataset into a training set, a validation set, and a test set. The training set is used to train the model and determine its parameters; the validation set is used to determine the training termination criteria and enhance the model's generalization capabilities; and the test set is used to verify the model's performance. The dataset division is shown in Table 1.

[0083] Table 1 Dataset division

[0084]

[0085] This example uses the same dataset to train a traditional multi-channel neural network quantile regression model (QRNN) and the proposed wind power generation probability prediction model (I-QRNN). The performance of the prediction models is verified by comparing the results of the two models. To fully describe the fluctuating characteristics of wind power, the present invention generates quantiles at 5%, 10%, ..., 85%, 90%, and 95%, using the 5% quantile as an interval, and then forms confidence intervals of 90%, 80%, ..., 20%, and 10%, respectively.

[0086] Figure 6The reliability metrics of the QRNN and I-QRNN models, obtained using the test set, are shown as a function of the look-ahead time, with 80% and 90% confidence intervals, to assess model reliability. It can be seen that the reliability of the I-QRNN model at 80% and 90% confidence levels for the first four hours is not significantly different from that of the QRNN model. Data source analysis shows that the accuracy of numerical weather forecasts has a certain impact on the reliability of forecast results. The shorter the look-ahead time, the higher the accuracy of the numerical weather forecast and the more reliable the forecast results. As the forecast timescale increases, the numerical weather forecast data becomes increasingly inaccurate, leading to a gradual deterioration in the reliability of the forecast results. On the other hand, when the input data is consistent, the reliability of the I-QRNN model is slightly higher than that of the QRNN model, demonstrating that the proposed improvements are effective and have a positive impact on forecast results. Experimental data demonstrate that the I-QRNN model has better forecasting performance and performs better in terms of reliability.

[0087] Figure 7 The graph plots the acuity of the QRNN and I-QRNN models as a function of the look-ahead time, under 90% and 80% confidence intervals. The horizontal axis shows the prediction timescale, and the vertical axis shows the acuity. For the same prediction model, the acuity at the 80% confidence interval is significantly higher than that at the 90% confidence interval, with the difference ranging between 5.8% and 12.2%. As is readily apparent, smaller confidence intervals have narrower widths, resulting in better acuity at the 80% confidence interval. On the other hand, the quantiles obtained by the I-QRNN approach have slightly higher acuity than the QRNN model, with the acuity of the I-QRNN model being approximately 4.5% lower. Although the QRNN model is essentially the same as the I-QRNN model, they differ significantly in their neural network and functional structures, demonstrating the effectiveness of the improvements proposed in this chapter. Furthermore, the graph shows that the acuity increases with increasing prediction time. This is because the accuracy of numerical weather forecasts deteriorates over longer timeframes, making wind power forecasting more difficult and leading to greater uncertainty. Overall, the research method proposed in this embodiment has good sensitivity.

[0088] Figure 8 and Figure 9 The changes in the 10%, 20%, ..., and 90% confidence intervals of the I-QRNN model and the QRNN model on November 12, 2016 are shown. The prediction time scale is 72 hours and the time resolution is 15 minutes. It can be seen that the observed values ​​of wind power are well surrounded by the confidence intervals, and the performance of the I-QRNN model is significantly better than that of the QRNN. Obviously, the low-level confidence intervals are well surrounded by the high-level confidence intervals, indicating that the proposed method effectively avoids the intersection of quantile curves. At the same time, Figure 9The above analysis results are verified. Based on the improved multi-channel neural network quantile regression wind power probability prediction model, a model with high learning ability and versatility is proposed.

[0089] Example 2

[0090] In this embodiment, a wind power prediction system based on an improved neural network quantile regression is disclosed, comprising:

[0091] A data acquisition module, used to obtain weather forecast data for multiple wind farms;

[0092] The wind power generation power prediction module is used to obtain the wind power generation power prediction results of each wind farm based on weather forecast data and a trained wind power generation power probability prediction model. The wind power generation power probability prediction model takes the weather forecast data of each wind farm as input and the wind power generation power prediction results as output. It is constructed using a neural network expression quantile regression algorithm, and the activation function of the model output layer uses a ramp function, and a smoothing function is introduced in the ramp function and the test function of the quantile regression algorithm.

[0093] Example 3

[0094] In this embodiment, an electronic device is disclosed, including a memory and a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, the steps described in the wind power prediction method based on improved neural network quantile regression disclosed in Example 1 are completed.

[0095] Example 4

[0096] In this embodiment, a computer-readable storage medium is disclosed for storing computer instructions. When the computer instructions are executed by a processor, the steps of the wind power prediction method based on improved neural network quantile regression disclosed in Example 1 are completed.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A wind power prediction method based on improved neural network quantile regression is characterized by: include: Obtain weather forecast data for multiple wind farms; Based on weather forecast data and a trained wind power generation probability prediction model, wind power generation prediction results for each wind farm are obtained. The wind power generation probability prediction model uses the weather forecast data for each wind farm as input and the wind power generation prediction results as output. It is constructed using a neural network expression quantile regression algorithm, and the activation function of the model output layer uses a ramp function. A smoothing function is introduced in the ramp function and the test function of the quantile regression algorithm. The neural network includes multiple input layers, each of which is connected to multiple hidden layers connected in sequence. The input of each input layer is the weather forecast data of a wind farm. The number of input layers is the same as the number of wind farms. The weather forecast data of each wind farm is input into the corresponding input layer, and then input into the hidden layer through the input layer for feature extraction. The features extracted by the hidden layer are output through the output layer. After introducing the smoothing function into the ramp function, the corrected ramp function is obtained: After introducing the smoothing function into the test function, the corrected test function is obtained: The loss function of the wind power generation probability prediction model is: in, is the ramp function, For what you want The quantile at the level, For what you want The quantile at the level, is the corrected test function, For what you want The quantile at the level, is the introduced smoothing function, is the loss function of the prediction model, is the number of samples, is the probability level, is the corrected test function, is the predicted value of wind power generation of the wind farm.

2. The wind power prediction method based on improved neural network quantile regression according to claim 1, characterized in that: Obtain weather forecast data and corresponding wind power generation data for multiple wind farms within a preset time period, and construct training and validation sets based on the acquired data; The constructed wind power generation power probability prediction model is trained using the training set, and the training termination condition of the wind power generation power prediction model is determined using the validation set. When the training is terminated, the trained wind power generation power probability prediction model is obtained.

3. The wind power prediction method based on improved neural network quantile regression according to claim 1, characterized in that: The gradient descent method is used to train the constructed wind power generation probability prediction model.

4. A wind power prediction system based on improved neural network quantile regression, which executes the wind power prediction method based on improved neural network quantile regression according to any one of claims 1 to 3, characterized in that: include: A data acquisition module, used to obtain weather forecast data for multiple wind farms; The wind power generation power prediction module is used to obtain the wind power generation power prediction results of each wind farm based on weather forecast data and a trained wind power generation power probability prediction model. The wind power generation power probability prediction model takes the weather forecast data of each wind farm as input and the wind power generation power prediction results as output. It is constructed using a neural network expression quantile regression algorithm, and the activation function of the model output layer uses a ramp function, and a smoothing function is introduced in the ramp function and the test function of the quantile regression algorithm.

5. An electronic device, characterized in that: The invention comprises a memory and a processor and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of the wind power prediction method based on improved neural network quantile regression according to any one of claims 1 to 3 are completed.

6. A computer-readable storage medium, characterized in that Used to store computer instructions, which, when executed by a processor, complete the steps of the wind power prediction method based on improved neural network quantile regression according to any one of claims 1 to 3.

Images