A wind power forecasting method based on fluctuation sequence classification correction

Abstract

The present invention discloses a wind power prediction method based on fluctuation sequence classification and correction, comprising: 1. using a benchmark model to predict the wind power benchmark value within the next N hours; 2. using a feature clustering method to divide the power fluctuation process, and exploring the correlation between meteorological forecast errors and model generalization errors under different fluctuation sequences from the perspective of output power; 3. using a CNN-LSTM time series model 1 to deduce power changes in future time periods for small fluctuation sequences with smooth output; and for non-small fluctuation sequences, combining a CNN-LSTM time series model 2 and a back-propagation neural network to interactively correct double-layer errors; and 4. recombining the benchmark power correction results according to the time series as the final wind power output. The present invention adopts a composite method combining time series analysis and feature learning to extract features from multiple dimensions to correct errors, and conforms to the actual error distribution law to ensure the model has good accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of short-term wind power forecasting, and in particular to a wind power forecasting method based on fluctuation sequence classification and correction. Background Art

[0002] Over the past decade, renewable energy, particularly wind power, has experienced rapid growth, gradually transitioning from an auxiliary power source to a primary power source. However, the large-scale integration of randomly fluctuating wind power into the power grid poses challenges to its safe and stable operation. Therefore, accurate wind power forecasting is crucial for improving wind power consumption and ensuring the economical and secure dispatch of power systems. Currently, short-term wind power forecasting methods can be categorized as physical and statistical. These methods rely heavily on the accuracy of numerical weather forecasts (NWPs) and historical training data. A single forecasting method can lead to significant deviations in prediction results, particularly for unexpected weather conditions. Numerical weather forecasting simulates actual meteorological processes by establishing a series of high-dimensional, nonlinear mathematical models. However, due to the instability of weather systems and the incompleteness of mathematical models, NWP data can only approximate future weather patterns for drastic atmospheric motions. Furthermore, a single forecasting method struggles to capture the patterns of wind power fluctuations, and models suffer from generalization errors when applied to unknown datasets. After outlier monitoring and adjustment of historical data of wind farms, the forecast error of wind power mainly comes from the insufficient accuracy of numerical weather forecast data and model generalization error, among which meteorological forecast error is the main source of forecast system error.

[0003] To address the above issues, some studies have used the regularity of the errors between the wind speed series in numerical weather forecasts and the actual wind speed series to correct the wind speed forecast through neural networks, reduce the errors in key meteorological characteristics, and improve the prediction accuracy of the prediction model. However, when it comes to correcting data deviations in weather forecasts, most studies often only select key meteorological factors such as wind speed and wind direction for correction, without taking into account the overall meteorological changes. The degree of data refinement needs to be improved.

[0004] Currently, comprehensive correction strategies and combined models applicable to multiple scenarios are becoming a research hotspot, offering new insights into wind power forecasting. Some studies have combined machine learning algorithms with benchmark forecasts to deduce overall meteorological patterns and correct benchmark values ​​point by point. However, these error correction steps fail to distinguish between meteorological forecast errors and model generalization errors, resulting in limited forecast accuracy during certain time periods. Therefore, combining real-time meteorological data with the error correction process of algorithmic models to dynamically correct power benchmark values ​​and improve model prediction accuracy during periods of rapid meteorological change is an urgent issue. Summary of the Invention

[0005] In order to overcome the shortcomings of the above-mentioned existing technologies, the present invention proposes a wind power prediction method based on fluctuation sequence classification correction, in order to explore the relationship between meteorological forecast errors and model generalization errors under different fluctuations from the perspective of power. A composite method combining time series analysis and feature learning is adopted to extract features from multiple dimensions to correct errors, thereby improving the prediction accuracy of the wind power model.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] The wind power prediction method based on fluctuation sequence classification correction of the present invention is characterized by being performed in the following steps:

[0008] Step 1: Select a sample set consisting of n historical samples of historical actual wind power data corresponding to l measured meteorological characteristics, l numerical weather forecast characteristics, and the measured meteorological characteristics, and normalize the sample set to obtain a measured meteorological characteristic set X with a dimension of n×l. n×l , the numerical weather forecast feature set Y with dimension n×l n×l and the actual power sequence P of n historical samples;

[0009] Step 2: Use the historical meteorological feature set X with dimension n×l n×l The actual power sequence P of n historical samples is used to train a back propagation neural network, and a benchmark model between meteorological data and fluctuating power is established;

[0010] Step 3: Based on the prediction results and prediction errors of the benchmark model, the prediction results are divided and the prediction errors are corrected using a data-driven strategy:

[0011] Step 3.1: Use equations (1), (2), and (3) to define the range of weather forecast error and model generalization error from the perspective of output power:

[0012] w nwp =f r -f nwp (1)

[0013] w m =f t -f r (2)

[0014] w=w nwp +w m (3)

[0015] In formula (1), formula (2), and formula (3), f r represents the predicted value of the benchmark model with measured meteorological characteristics as input, f nwprepresents the predicted value of the benchmark model with numerical weather forecast characteristics as input, f t Indicates the actual wind power value, w nwp represents the weather forecast error, w m represents the model generalization error, and w represents the overall error of the benchmark model prediction result, that is, the deviation between the predicted value of the benchmark model with the normalized historical numerical weather forecast characteristics as input and the actual value of wind power at the corresponding moment;

[0016] Step 3.2: Use the feature clustering method to divide the actual power sequence P of n historical samples and obtain the fluctuation sequence. Then use equations (4) and (5) to identify the four power fluctuation processes of the fluctuation sequence:

[0017]

[0018]

[0019] In formula (4) and formula (5), P max_1 ,P max_2 are the first two peaks after removing all peaks of meteorological disturbances in the fluctuation process and sorting them in descending order; G is the fluctuation category, P is the fluctuation category, max is the maximum peak value in the fluctuation process, K is the peak ratio, ε1 is the discrimination threshold for low power output, ε2 is the discrimination threshold for high power output, and K′ is the threshold for the peak ratio;

[0020] Step 3.3: Extract small fluctuation processes from the fluctuation sequence and combine them into a small fluctuation sequence according to the time sequence. The remaining fluctuation processes are combined into a non-small fluctuation sequence according to the time sequence.

[0021] Step 3.4, selecting the wind power forecast value output by the benchmark model after the numerical weather forecast characteristics within Z hours before any moment T in the historical period where the small fluctuation sequence is located, the normalized wind speed sequence in the numerical weather forecast characteristics within Z hours before any moment T in the historical period where the small fluctuation sequence is located, and the overall error of the benchmark model prediction result within Z hours before any moment T in the historical period where the small fluctuation sequence is located as the three inputs of the CNN-LSTM time series model 1, and selecting the overall error of the benchmark model prediction result at any moment T in the historical period where the small fluctuation sequence is located as the output of the CNN-LSTM time series model 1, thereby using the small fluctuation sequence and its corresponding numerical weather forecast characteristics in the sample set to train the CNN-LSTM time series model 1, and obtaining the trained CNN-LSTM time series model 1;

[0022] Step 3.5, selecting the wind power forecast value output by the benchmark model after the numerical weather forecast characteristics within Z′ hours before any moment T′ in the historical period where the non-small fluctuation sequence is located, the model generalization error within Z′ hours before any moment T′ in the historical period where the non-small fluctuation sequence is located, and the weather forecast error within Z′ hours before any moment T′ in the historical period where the non-small fluctuation sequence is located as the three inputs of the CNN-LSTM time series model 2, and selecting the weather forecast error at any moment T′ in the historical period where the non-small fluctuation sequence is located as the output of the CNN-LSTM time series model 2, thereby using the non-small fluctuation sequence and its corresponding numerical weather forecast characteristics in the sample set to train the CNN-LSTM time series model 2, and obtain the trained CNN-LSTM time series model 2;

[0023] Step 3.6: Use formula (6) to extract the numerical weather forecast feature set V of the historical period where the non-small fluctuation series is located. m×l The features that are strongly correlated with the model generalization error y in the historical period of the non-small fluctuation series are selected:

[0024]

[0025] In formula (6), r s is the numerical weather forecast feature set V for the historical period of the non-small fluctuation series m×l Pearson correlation coefficient between the sth feature in and the model generalization error y in the historical period of the non-small fluctuation series, s = 1,…,l; The numerical weather forecast feature set V represents the historical period of the non-small fluctuation series m×l The sth feature V s The average value of represents the average value of the model generalization error y in the historical period where the non-small fluctuation sequence is located; m is the number of samples in the historical period where the non-small fluctuation sequence is located; V ps The numerical weather forecast feature set V represents the historical period of the non-small fluctuation series m×l The pth sample value of the sth feature in y p Represents the pth sample value of the model generalization error y in the historical period of the non-small fluctuation series;

[0026] Step 3.7: The absolute value of the Pearson correlation coefficient | r s | Sort in descending order and select the first q features as the numerical weather forecast features that have a strong correlation with the generalization error;

[0027] Step 3.8, selecting the wind power forecast value output by the benchmark model after the numerical weather forecast characteristics in the historical period of the non-small fluctuation sequence, the meteorological forecast error in the historical period of the non-small fluctuation sequence, and the numerical weather forecast characteristics that have a strong correlation with the generalization error as three inputs of the back propagation neural network, and selecting the model generalization error in the historical period of the non-small fluctuation sequence as the output of the back propagation neural network, thereby training the back propagation neural network using the non-small fluctuation sequence and its corresponding numerical weather forecast characteristics in the sample set, and obtaining a trained back propagation neural network;

[0028] Step 4: Inputting the numerical weather forecast feature Y′ of the current forecast period into the benchmark model, and predicting the wind power benchmark value of the current forecast period;

[0029] Step 5: Divide the wind power benchmark value of the current forecast period into a fluctuation sequence and match it with the corresponding correction model to correct the forecast error of the benchmark model:

[0030] Step 5.1: Divide the wind power benchmark value of the current forecast period to obtain the fluctuation sequence of the current forecast period, and extract the small fluctuation sequence and non-small fluctuation sequence of the current forecast period;

[0031] Step 5.2: Use the trained CNN-LSTM time series model 1 to perform rolling correction on the small fluctuation sequence:

[0032] Step 5.2.1: Define the single prediction range of the small fluctuation series as [t1,t z ], define the variable t j ∈[t1,t z ], j = 1,…, z;

[0033] Step 5.2.2, with t j The wind power forecast value output by the benchmark model after the numerical weather forecast characteristics within Z hours before time t j Normalized wind speed series and t in numerical weather forecast characteristics within Z hours before time j The overall error of the baseline model prediction results within Z hours before the moment is used as the three inputs of the trained CNN-LSTM time series model 1 to deduce the output t j The overall error w of the baseline model prediction result at time j ;

[0034] Step 5.2.3: After assigning j+1 to j, determine whether j>z. If so, the error correction within the single prediction range is complete, and proceed to step 5.3. Otherwise, return to step 5.2.2.

[0035] Step 5.3: Use the trained CNN-LSTM time series model 2 and the back propagation neural network to interactively correct the non-small fluctuation sequence:

[0036] Step 5.3.1: Define the single prediction range of the non-small fluctuation series as [t1,t n ], variable t i ∈[t1,t n ], i=1,…,n;

[0037] Step 5.3.2, with t i The wind power prediction value, model generalization error and weather forecast error output by the benchmark model after the numerical weather forecast characteristics within Z′ hours before the time are respectively used as the three inputs of the trained CNN-LSTM time series model 2, thereby deducing the output t i Weather forecast error at time w nwp_i ;

[0038] Step 5.3.3, with t i The wind power forecast value output by the benchmark model after the numerical weather forecast characteristics at time t i Weather forecast error at time w nwp_i and t i The numerical weather forecast features that are strongly correlated with the generalization error at all times are used as the input of the trained back propagation neural network, thereby outputting t i The model generalization error w at time m_i ;

[0039] Step 5.3.4: After assigning i+1 to i, determine whether i>n holds. If so, error correction within the single prediction range is complete, and proceed to step 5.4. Otherwise, return to step 5.3.2.

[0040] Step 5.4: Superimpose the wind power reference value for the current forecast period and the overall error as the benchmark power correction result for the small fluctuation sequence; superimpose the wind power reference value for the current forecast period, the meteorological forecast error, and the model generalization error as the benchmark power correction result for the non-small fluctuation sequence;

[0041] Step 5.5: The reference power correction results of the small fluctuation sequence and the non-small fluctuation sequence are recombined according to the time sequence as the final wind power output.

[0042] Compared with the prior art, the beneficial effects of the present invention are embodied in:

[0043] 1 The present invention divides the fluctuation sequence based on the wind power benchmark value, which is conducive to error correction. It comprehensively analyzes the error conditions under different fluctuation sequences, identifies the power small fluctuation process, and reorganizes the corresponding power fluctuation into small fluctuation sequence and non-small fluctuation sequence according to the time sequence. Different fluctuation sequences adopt corresponding correction strategies, which improves the correction accuracy of wind power and avoids the situation where the overall correction model is difficult to fit the error distribution law and the prediction results oscillate around the actual power.

[0044] 2 Aiming at the sequence of small fluctuations with smooth output, the present invention adopts the CNN-LSTM time series model to analyze historical similar scenarios for correction, selects key features from historical wind power information, makes full use of adjacent time information to track power fluctuations, and deduce the changing trend of future wind power output.

[0045] 3. Aiming at non-small fluctuation sequences with large fluctuations, the present invention implements the data-driven concept, combines multi-scale features with time series deduction and feature-based learning to dynamically correct the benchmark results, and uses real-time feedback of the correction results to fit the actual error distribution law, thereby improving the model's prediction effect on wind power fluctuation periods. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a schematic diagram of the wind power prediction process for the current forecast period;

[0047] Figure 2 Schematic diagram of the two-layer error interactive correction process within a single prediction range;

[0048] Figure 3 This is a comparison chart of the prediction results of different methods during some periods of summer;

[0049] Figure 4 This is a comparison chart of the prediction results of different methods during some periods of winter. DETAILED DESCRIPTION

[0050] In this embodiment, a wind power prediction method based on fluctuation sequence classification and correction combines multi-scale feature extraction to explore the error correlation under different fluctuation sequences. Especially during periods of sharp fluctuations, the method uses real-time error feedback during the interactive correction process to adapt to the actual error distribution pattern and dynamically correct the power reference value at the moment of large fluctuations to improve the accuracy and applicability of the prediction method. Specifically, the prediction method is carried out in the following steps:

[0051] Step 1: Select a sample set consisting of n historical samples of measured meteorological characteristics, l numerical weather forecast characteristics, and historical actual wind power data corresponding to the measured meteorological characteristics, and normalize the sample set to obtain a measured meteorological feature set X with a dimension of n×l. n×l , the numerical weather forecast feature set Y with dimension n×ln×l and the actual power sequence P of n historical samples;

[0052] Step 2: Use the historical meteorological feature set X with dimension n×l n×l The actual power sequence P of n historical samples is used to train a back propagation neural network, and a benchmark model between meteorological data and fluctuating power is established;

[0053] Wind power fluctuations involve a variety of meteorological factors, but too many input features can make model training more difficult. Therefore, for the baseline prediction model, a back-propagation neural network is trained using historically measured key meteorological data as input and historical actual wind power as output, establishing a mapping between meteorological data and fluctuating power.

[0054] Step 3: Based on the prediction results and prediction errors of the baseline model, the prediction results are divided and the prediction errors are corrected using a data-driven strategy:

[0055] Step 3.1: Use equations (1), (2), and (3) to define the range of weather forecast error and model generalization error from the perspective of output power:

[0056] w nwp =f r -f nwp (1)

[0057] w m =f t -f r (2)

[0058] w=w nwp +w m (3)

[0059] In formula (1), formula (2), and formula (3), f r represents the predicted value of the benchmark model with measured meteorological characteristics as input, f nwp represents the predicted value of the benchmark model with numerical weather forecast characteristics as input, f t Indicates the actual wind power value, w nwp represents the weather forecast error, w m represents the model generalization error, and w represents the overall error of the benchmark model prediction result, that is, the deviation between the predicted value of the benchmark model with the normalized historical numerical weather forecast characteristics as input and the actual value of wind power at the corresponding moment, which includes two parts: meteorological forecast error and model generalization error;

[0060] Step 3.2: Use the feature clustering method to divide the actual power sequence P of n historical samples and obtain the fluctuation sequence. Then use equations (4) and (5) to identify the four power fluctuation processes of the fluctuation sequence:

[0061]

[0062]

[0063] In formula (4) and formula (5), P max_1 ,P max_2 are the first two peaks after removing all peaks of meteorological disturbances in the fluctuation process and sorting them in descending order; G is the fluctuation category, P is the fluctuation category, max is the maximum peak value in the fluctuation process, K is the peak ratio, ε1 is the discrimination threshold for low power output, ε2 is the discrimination threshold for high power output, and K′ is the threshold for the peak ratio;

[0064] During small fluctuations, wind power output is at a low level, meteorological conditions do not change much, and numerical weather forecast data can basically predict meteorological fluctuations, so the benchmark model has a certain degree of prediction accuracy; for medium and large fluctuations, wind power output gradually increases, meteorological fluctuations gradually become more intense, there is a certain gap between numerical weather forecast data and the actual weather, and the model prediction results will have a large deviation; compared with medium and large fluctuations, multi-peak oscillation processes have a longer overall duration, and the process contains multiple similar fluctuation segments. The meteorological conditions change complexly, and the benchmark model is difficult to track power fluctuations. Four types of fluctuation processes are divided, and the error conditions under different fluctuation processes are studied, and targeted strategies are adopted to correct the corresponding errors. In addition, the peak sizes within the fluctuation process are sorted in descending order, and adjacent peaks in a very short time are regarded as disturbances. The average of the adjacent peaks is taken as its peak output, which removes the influence of meteorological disturbances. The two largest peaks in the entire process are recorded as P max_1 ,P max_2 , and P max is the maximum peak value in the fluctuation process without disturbance treatment, so P max_1 The maximum value P of the peak value in the fluctuation process max There is a difference.

[0065] Step 3.3: Extract small fluctuation processes from the fluctuation sequence and combine them into a small fluctuation sequence according to the time sequence. The remaining fluctuation processes are combined into a non-small fluctuation sequence according to the time sequence.

[0066] In small fluctuation sequences, the model can fit the data well, and the error is concentrated in the meteorological forecast link. For non-small fluctuation sequences, the double-layer error cannot be ignored, and the correction of meteorological error and generalization error is considered. At the same time, in order to maintain the integrity of the original sequence as much as possible, the information of adjacent moments is fully utilized to track power fluctuations, and only the small power fluctuation process is identified. The corresponding power fluctuations are reorganized into small fluctuation sequences and non-small fluctuation sequences according to the time series. The impact of meteorological forecast deviations in similar historical scenarios on the power baseline value of small fluctuation sequences is analyzed. For non-small fluctuation sequences, the correlation between the double-layer errors is further explored to improve the accuracy of the model in moments of large fluctuations.

[0067] Step 3.4: Select the wind power forecast value output by the benchmark model after the numerical weather forecast characteristics within Z hours before any moment T in the historical period where the small fluctuation sequence is located, the normalized wind speed sequence in the numerical weather forecast characteristics within Z hours before any moment T in the historical period where the small fluctuation sequence is located, and the overall error of the benchmark model prediction result within Z hours before any moment T in the historical period where the small fluctuation sequence is located as the three inputs of the CNN-LSTM time series model 1, and select the overall error of the benchmark model prediction result at any moment T in the historical period where the small fluctuation sequence is located as the output of the CNN-LSTM time series model 1, so as to train the CNN-LSTM time series model 1 using the small fluctuation sequence and its corresponding numerical weather forecast characteristics in the sample set, and obtain the trained CNN-LSTM time series model 1;

[0068] Time series models analyze the underlying features and connections of historical sequences to predict future data changes. The CNN-LSTM model connects multiple convolutional and pooling layers to extract internal features from the data and use them as neural network input for time series prediction. The CNN model uses different convolution kernels in the convolutional layers to gradually explore the inherent connections between data. The pooling layer extracts important features from these layers, simplifying the feature dimensions while enhancing the model's predictive capabilities. The LSTM network then filters relevant information and uses historical data to predict future trends.

[0069] Step 3.5: Select the wind power forecast value output by the benchmark model after the numerical weather forecast characteristics within Z′ hours before any moment T′ in the historical period where the non-small fluctuation sequence is located, the model generalization error within Z′ hours before any moment T′ in the historical period where the non-small fluctuation sequence is located, and the weather forecast error within Z′ hours before any moment T′ in the historical period where the non-small fluctuation sequence is located as the three inputs of the CNN-LSTM time series model 2, and select the weather forecast error at any moment T′ in the historical period where the non-small fluctuation sequence is located as the output of the CNN-LSTM time series model 2, so as to train the CNN-LSTM time series model 2 using the non-small fluctuation sequence and its corresponding numerical weather forecast characteristics in the sample set, and obtain the trained CNN-LSTM time series model 2;

[0070] The meteorological error correction results, combined with the power reference value, are fed into the generalization error correction model as real-time features. This serves to identify fluctuation types, allowing for rapid error correction while avoiding model classification. Furthermore, while meteorological fluctuation patterns are difficult to grasp, real-time meteorological variations are significant during power peaks and troughs. The positive or negative sign of the generalization error not only marks the output moment, but its magnitude indirectly reflects the severity of the real-time meteorological fluctuation, integrating both marker and quantitative features. Furthermore, as the prediction step increases over time, the time series dependency decreases, and the prediction error gradually increases. By feeding realistic generalization errors into the time series model and selecting real-time features as input, error accumulation can be effectively reduced, enhancing the time series model's ability to capture the impact of future meteorological forecast deviations on the power reference value. Therefore, for non-small fluctuation series, the two-layer error exhibits a certain degree of correlation. Using the correction results of the two-layer error for interactive feedback improves the model's prediction accuracy during periods of significant meteorological fluctuations.

[0071] Step 3.6: Use formula (6) to extract the numerical weather forecast feature set V of the historical period where the non-small fluctuation series is located. m×l The features that are strongly correlated with the model generalization error y in the historical period of the non-small fluctuation series are selected:

[0072]

[0073] In formula (6), r s is the numerical weather forecast feature set V for the historical period of the non-small fluctuation series m×l Pearson correlation coefficient between the sth feature in and the model generalization error y in the historical period of the non-small fluctuation series, s = 1,…,l; The numerical weather forecast feature set V represents the historical period of the non-small fluctuation series m×l The sth feature V s The average value of represents the average value of the model generalization error y in the historical period where the non-small fluctuation sequence is located; m is the number of samples in the historical period where the non-small fluctuation sequence is located; V ps The numerical weather forecast feature set V represents the historical period of the non-small fluctuation series m×l The pth sample value of the sth feature in y p Represents the pth sample value of the model generalization error y in the historical period of the non-small fluctuation series;

[0074] Step 3.7: The absolute value of the Pearson correlation coefficient | r s | Sort in descending order and select the first q features as the numerical weather forecast features that have a strong correlation with the generalization error;

[0075] Because numerical weather forecast features include numerous meteorological characteristics, the correlation between model generalization error and these characteristics varies. Selecting highly correlated features can simplify the model structure and accelerate model training. Furthermore, the Pearson correlation coefficient can be used to measure the correlation between variables. Its value range is [-1, 1]. When the correlation coefficient is greater than zero, the two variables are positively correlated, while when the correlation coefficient is less than zero, the correlation is negative. The larger the absolute value of the correlation coefficient, the higher the correlation between the variables. The Pearson correlation coefficient can be used to select and verify model input features.

[0076] Step 3.8: Select the wind power forecast value output by the benchmark model after the numerical weather forecast characteristics in the historical period of the non-small fluctuation sequence, the meteorological forecast error in the historical period of the non-small fluctuation sequence, and the numerical weather forecast characteristics that have a strong correlation with the generalization error as the three inputs of the back propagation neural network, and select the model generalization error in the historical period of the non-small fluctuation sequence as the output of the back propagation neural network, thereby using the non-small fluctuation sequence and its corresponding numerical weather forecast characteristics in the sample set to train the back propagation neural network, and obtain a trained back propagation neural network;

[0077] Step 4: Input the numerical weather forecast feature Y′ of the current forecast period into the benchmark model to predict the wind power benchmark value of the current forecast period;

[0078] When predicting the wind power in the current forecast period, since the measured meteorological characteristics are unknown, the numerical weather forecast characteristics are used as the model input to determine the power reference value. The wind power forecast process diagram for the current forecast period is as follows: Figure 1 shown.

[0079] Step 5: Divide the wind power benchmark value of the current forecast period into a fluctuation sequence and match it with the corresponding correction model to correct the forecast error of the benchmark model:

[0080] Step 5.1: Divide the wind power benchmark value of the current forecast period to obtain the fluctuation sequence of the current forecast period, and extract the small fluctuation sequence and non-small fluctuation sequence of the current forecast period;

[0081] Step 5.2: Use the trained CNN-LSTM time series model 1 to perform rolling correction on the small fluctuation sequence:

[0082] Step 5.2.1: Define the single prediction range of the small fluctuation series as [t1,t z ], variable t j ∈[t1,t z ], j = 1,…, z;

[0083] Step 5.2.2, with t j The wind power forecast value output by the benchmark model after the numerical weather forecast characteristics within Z hours before time t j Normalized wind speed series and t in numerical weather forecast characteristics within Z hours before time j The overall error of the baseline model prediction results within Z hours before the moment is used as the three inputs of the trained CNN-LSTM time series model 1 to deduce the output t j The overall error w of the baseline model prediction result at time j ;

[0084] Step 5.2.3: After assigning j+1 to j, determine whether j>z. If so, the error correction within the single prediction range is complete, and proceed to step 5.3. Otherwise, return to step 5.2.2.

[0085] During the training process of CNN-LSTM time series model 1, the overall error of the benchmark model prediction result, the wind power prediction value output by the numerical weather forecast feature after passing through the benchmark model, and the normalized wind speed sequence in the numerical weather forecast feature can be obtained from historical data. Therefore, the model is trained in a single-point output manner. However, within the single prediction range of the current forecast period, the overall error of the actual benchmark model prediction result cannot be obtained. Only a rolling correction method can be adopted. The overall error of the benchmark model prediction result at the current j-th forecast moment is used to replace the overall error of the actual benchmark model prediction result. The input is input into the trained CNN-LSTM time series model 1 at the j+1-th forecast moment to perform error correction of the small fluctuation sequence. The remaining model inputs can be obtained through the numerical weather forecast features of the current period, thereby completing the error correction within the single prediction range.

[0086] Step 5.3: Use the trained CNN-LSTM time series model 2 and the back propagation neural network to interactively correct the non-small fluctuation sequence:

[0087] Step 5.3.1: Define the single prediction range of the non-small fluctuation series as [t1,t n ], variable t i ∈[t1,t n ], i=1,…,n;

[0088] Step 5.3.2, with t i The wind power forecast value, model generalization error, and weather forecast error output by the benchmark model within the hour before the time Z′ are used as the three inputs of the trained CNN-LSTM time series model 2, thereby deducing the output t i Weather forecast error at time w nwp_i ;

[0089] Step 5.3.3, with t i The wind power forecast value output by the benchmark model after the numerical weather forecast characteristics at time t i Weather forecast error at time w nwp_i and t i The numerical weather forecast features that are strongly correlated with the generalization error at all times are used as the input of the trained back propagation neural network, thereby outputting t i The model generalization error w at time m_i ;

[0090] Step 5.3.4: After assigning i+1 to i, determine whether i>n holds. If so, error correction within the single prediction range is complete, and proceed to step 5.4. Otherwise, return to step 5.3.2.

[0091] During the training process of CNN-LSTM time series model 2, the wind power prediction value, model generalization error and meteorological forecast error output by the numerical weather forecast feature after passing through the benchmark model can be obtained from historical data, so the model is trained in a single-point output manner; during the training process of the back propagation neural network, the wind power prediction value, meteorological forecast error and numerical weather forecast features with a strong correlation with the generalization error output by the numerical weather forecast feature after passing through the benchmark model can be obtained from historical data, so the model is trained in a feature learning-based manner; and within the single prediction range of the current forecast period, since the model generalization error and meteorological forecast error cannot be obtained, the only way is to use the interactive correction method to output t i The weather forecast error at time t i The weather forecast error at the moment is used as the input of the back propagation neural network, and the output is t i The generalization error of the model at the moment is iterated repeatedly until the error correction within the single prediction range is completed. The flow chart of the interactive correction of non-small fluctuation series within the single prediction range is as follows Figure 2 As shown in Figure 2, due to the correlation between the two-layer errors, the correction results of the two-layer errors are fed back to the correction model in real time to cope with the large fluctuations of meteorological factors.

[0092] Step 5.4: Superimpose the wind power reference value for the current forecast period and the overall error as the benchmark power correction result for the small fluctuation sequence; superimpose the wind power reference value for the current forecast period, the meteorological forecast error, and the model generalization error as the benchmark power correction result for the non-small fluctuation sequence;

[0093] Step 5.5: The reference power correction results of the small fluctuation sequence and the non-small fluctuation sequence are recombined according to the time sequence as the final wind power output.

[0094] Unexpected weather conditions pose considerable challenges to both weather forecasting and wind power prediction. Therefore, when dividing the data samples, summer and winter are selected as typical seasons. To verify the superiority of the method proposed in this paper, the following four methods are selected and compared with actual power. The comparison results in some periods of summer and winter are as follows: Figure 3 , Figure 4 Method 1: The numerical weather forecast characteristics of the current forecast period are used as the input of the benchmark model to predict the power benchmark value of the corresponding period; Method 2: Without dividing the fluctuation sequence, the double-layer error interactive correction method is used to correct the complete sequence, and the correction result is used as the wind power output; Method 3: Without dividing the fluctuation sequence, the CNN-LSTM model is used to correct the complete sequence, and the correction result is used as the wind power output; Method 4: The fluctuation sequence is divided, the CNN-LSTM model is used to correct the small fluctuation sequence, and the double-layer error interactive correction method is used to correct the non-small fluctuation sequence. Finally, the correction result is reorganized according to the time series as the wind power output.

[0095] In summary, the present invention divides the power sequence based on the power reference value, matches different fluctuation sequences with corresponding error correction strategies, and improves the correction accuracy of the model. Among them, the time series model CNN-LSTM is used to extract the intrinsic characteristics of the historical sequence, and analyzes historical similar scenarios to deduce the small fluctuation sequence error; for non-small fluctuation sequences with high double-layer error coupling, combined with multi-scale feature extraction, in the interactive correction process, the correction results are added to the correction model as new feature dimensions, and the error change trend is jointly deduced, avoiding the complexity of correcting meteorological errors one by one. At the same time, for scenarios where real-time features are difficult to extract, the correlation between meteorological errors and model generalization errors is explored, and real-time feedback is adopted to improve the prediction effect of the model when meteorological factors fluctuate greatly.

Claims

1. A wind power prediction method based on fluctuation sequence classification correction, characterized by Proceed as follows: Step 1: Select a sample set consisting of n historical samples of historical actual wind power data corresponding to l measured meteorological characteristics, l numerical weather forecast characteristics, and the measured meteorological characteristics, and normalize the sample set to obtain a measured meteorological characteristic set X with a dimension of n×l. n×l , the numerical weather forecast feature set Y with dimension n×l n×l and the actual power sequence P of n historical samples; Step 2: Use the historical meteorological feature set X with dimension n×l n×l The actual power sequence P of n historical samples is used to train a back propagation neural network, and a benchmark model between meteorological data and fluctuating power is established; Step 3: Based on the prediction results and prediction errors of the benchmark model, the prediction results are divided and the prediction errors are corrected using a data-driven strategy: Step 3.1: Use equations (1), (2), and (3) to define the range of weather forecast error and model generalization error from the perspective of output power: w nwp =f r -f nwp (1) w m =f t -f r (2) w=w nwp +in m (3) In formula (1), formula (2), and formula (3), f r represents the predicted value of the benchmark model with measured meteorological characteristics as input, f nwp represents the predicted value of the benchmark model with numerical weather forecast characteristics as input, f t Indicates the actual wind power value, w nwp represents the weather forecast error, w m represents the model generalization error, and w represents the overall error of the benchmark model prediction result, that is, the deviation between the predicted value of the benchmark model with the normalized historical numerical weather forecast characteristics as input and the actual value of wind power at the corresponding moment; Step 3.2: Use the feature clustering method to divide the actual power sequence P of n historical samples and obtain the fluctuation sequence. Then use equations (4) and (5) to identify the four power fluctuation processes of the fluctuation sequence: In formula (4) and formula (5), P max_1 ,P max_2 are the first two peaks after removing all peaks of meteorological disturbances in the fluctuation process and sorting them in descending order; G is the fluctuation category, P is the fluctuation category, max is the maximum peak value in the fluctuation process, K is the peak ratio, ε1 is the discrimination threshold for low power output, ε2 is the discrimination threshold for high power output, and K′ is the threshold for the peak ratio; Step 3.3: Extract small fluctuation processes from the fluctuation sequence and combine them into a small fluctuation sequence according to the time sequence. The remaining fluctuation processes are combined into a non-small fluctuation sequence according to the time sequence. Step 3.4, selecting the wind power forecast value output by the benchmark model after the numerical weather forecast characteristics within Z hours before any moment T in the historical period where the small fluctuation sequence is located, the normalized wind speed sequence in the numerical weather forecast characteristics within Z hours before any moment T in the historical period where the small fluctuation sequence is located, and the overall error of the benchmark model prediction result within Z hours before any moment T in the historical period where the small fluctuation sequence is located as the three inputs of the CNN-LSTM time series model 1, and selecting the overall error of the benchmark model prediction result at any moment T in the historical period where the small fluctuation sequence is located as the output of the CNN-LSTM time series model 1, thereby using the small fluctuation sequence and its corresponding numerical weather forecast characteristics in the sample set to train the CNN-LSTM time series model 1, and obtaining the trained CNN-LSTM time series model 1; Step 3.5, selecting the wind power forecast value output by the benchmark model after the numerical weather forecast characteristics within Z′ hours before any moment T′ in the historical period where the non-small fluctuation sequence is located, the model generalization error within Z′ hours before any moment T′ in the historical period where the non-small fluctuation sequence is located, and the weather forecast error within Z′ hours before any moment T′ in the historical period where the non-small fluctuation sequence is located as the three inputs of the CNN-LSTM time series model 2, and selecting the weather forecast error at any moment T′ in the historical period where the non-small fluctuation sequence is located as the output of the CNN-LSTM time series model 2, thereby using the non-small fluctuation sequence and its corresponding numerical weather forecast characteristics in the sample set to train the CNN-LSTM time series model 2, and obtain the trained CNN-LSTM time series model 2; Step 3.6: Use formula (6) to extract the numerical weather forecast feature set V of the historical period where the non-small fluctuation series is located. m×l The features that are strongly correlated with the model generalization error y in the historical period of the non-small fluctuation series are selected: In formula (6), r s is the numerical weather forecast feature set V for the historical period of the non-small fluctuation series m×l Pearson correlation coefficient between the sth feature in and the model generalization error y in the historical period of the non-small fluctuation series, s = 1,…,l; The numerical weather forecast feature set V represents the historical period of the non-small fluctuation series m×l The sth feature V s The average value of represents the average value of the model generalization error y in the historical period where the non-small fluctuation sequence is located; m is the number of samples in the historical period where the non-small fluctuation sequence is located; V ps The numerical weather forecast feature set V represents the historical period of the non-small fluctuation series m×l The pth sample value of the sth feature in y p Represents the pth sample value of the model generalization error y in the historical period of the non-small fluctuation series; Step 3.7: The absolute value of the Pearson correlation coefficient | r s | Sort in descending order and select the first q features as the numerical weather forecast features that have a strong correlation with the generalization error; Step 3.8, selecting the wind power forecast value output by the benchmark model after the numerical weather forecast characteristics in the historical period of the non-small fluctuation sequence, the meteorological forecast error in the historical period of the non-small fluctuation sequence, and the numerical weather forecast characteristics that have a strong correlation with the generalization error as three inputs of the back propagation neural network, and selecting the model generalization error in the historical period of the non-small fluctuation sequence as the output of the back propagation neural network, thereby training the back propagation neural network using the non-small fluctuation sequence and its corresponding numerical weather forecast characteristics in the sample set, and obtaining a trained back propagation neural network; Step 4: Inputting the numerical weather forecast feature Y′ of the current forecast period into the benchmark model, and predicting the wind power benchmark value of the current forecast period; Step 5: Divide the wind power benchmark value of the current forecast period into a fluctuation sequence and match it with the corresponding correction model to correct the forecast error of the benchmark model: Step 5.1: Divide the wind power benchmark value of the current forecast period to obtain the fluctuation sequence of the current forecast period, and extract the small fluctuation sequence and non-small fluctuation sequence of the current forecast period; Step 5.2: Use the trained CNN-LSTM time series model 1 to perform rolling correction on the small fluctuation sequence: Step 5.2.1: Define the single prediction range of the small fluctuation series as [t1,t z ], define the variable t j ∈[t1,t z ], j = 1,…, z; Step 5.2.2, with t j The wind power forecast value output by the benchmark model after the numerical weather forecast characteristics within Z hours before time t j Normalized wind speed series and t in numerical weather forecast characteristics within Z hours before time j The overall error of the baseline model prediction results within Z hours before the moment is used as the three inputs of the trained CNN-LSTM time series model 1 to deduce the output t j The overall error w of the baseline model prediction result at time j ; Step 5.2.3: After assigning j+1 to j, determine whether j>z. If so, the error correction within the single prediction range is complete, and proceed to step 5.

3. Otherwise, return to step 5.2.

2. Step 5.3: Use the trained CNN-LSTM time series model 2 and the back propagation neural network to interactively correct the non-small fluctuation sequence: Step 5.3.1: Define the single prediction range of the non-small fluctuation series as [t1,t n ], variable t i ∈[t1,t n ], i=1,…,n; Step 5.3.2, with t i The wind power prediction value, model generalization error and weather forecast error output by the benchmark model after the numerical weather forecast characteristics within Z′ hours before the time are respectively used as the three inputs of the trained CNN-LSTM time series model 2, thereby deducing the output t i Weather forecast error at time w nwp_i ; Step 5.3.3, with t i The wind power forecast value output by the benchmark model after the numerical weather forecast characteristics at time t i Weather forecast error at time w nwp_i and t i The numerical weather forecast features that are strongly correlated with the generalization error at all times are used as the input of the trained back propagation neural network, thereby outputting t i The model generalization error w at time m_i ; Step 5.3.4: After assigning i+1 to i, determine whether i>n holds. If so, error correction within the single prediction range is complete, and proceed to step 5.

4. Otherwise, return to step 5.3.

2. Step 5.4: Superimpose the wind power reference value for the current forecast period and the overall error as the benchmark power correction result for the small fluctuation sequence; superimpose the wind power reference value for the current forecast period, the meteorological forecast error, and the model generalization error as the benchmark power correction result for the non-small fluctuation sequence; Step 5.5: The reference power correction results of the small fluctuation sequence and the non-small fluctuation sequence are recombined according to the time sequence as the final wind power output.

Images