Fan power interval prediction method based on multistage decomposition and nuclear density quantile

Abstract

The invention provides a multistage decomposition and nuclear density quantile-based fan power interval prediction method, which comprises the following steps of: firstly, acquiring a fan power data set, and screening related characteristic data; the method comprises the following steps: preprocessing data, dividing the data into a training set and a test set, and constructing an input feature vector according to the obtained data; secondly, variational mode decomposition is carried out on the sequence, and key parameters are optimized; thirdly, a PMGRU prediction model used for fan power interval prediction is constructed; and then, training the PMGRU prediction model by using the training set to obtain a trained model. And finally, running the trained interval prediction model on the test set to obtain a prediction result. According to the method, internal and external interference factors influencing the output power of the fan are fully considered, and the problem that the time correlation cannot be clearly captured under the influence of multiple factors is solved.

Description

Technical Field

[0001] The present invention belongs to the field of wind power prediction, and in particular relates to a wind turbine power interval prediction method based on multi-level decomposition and kernel density quantiles. Background Art

[0002] In recent years, the proportion of renewable energy has continued to grow globally. Wind energy, due to its pollution-free and renewable nature, has attracted considerable attention, and the scale of wind power generation has continued to expand. However, the combined impact of social activities, meteorological changes, and equipment conditions has led to significant volatility, non-stationarity, and uncertainty, posing unprecedented challenges to power systems. Accurate and timely forecasting of wind power generation is a prerequisite for scientific decision-making in power systems. It helps grid operators plan and adjust the safe and stable operation of the power system and manage demand-side energy distribution.

[0003] Wind turbine power forecasting is influenced by numerous factors, including meteorological variations, seasonal fluctuations, and hardware. The nonlinear relationships between these factors lead to complex fluctuations and non-stationarity in wind turbine output power, making it difficult for prediction models to capture the coupled correlations and temporal characteristics between features. With the rapid development of deep learning, it has also been rapidly applied to wind power interval forecasting, showing great promise in this field. Various LSTM and TCN-based models have been used to address issues such as reduced temporal continuity and local extremes that arise from this non-stationarity. Therefore, research addressing the difficulties in identifying non-stationary features and extracting temporal information is crucial for building accurate interval prediction models. Summary of the Invention

[0004] The present invention aims to overcome the uncertainty caused by the non-stationarity and volatility of wind turbine power prediction and the difficulty in extracting deep features. Furthermore, a single point prediction model cannot guarantee the accuracy and stability of the prediction. Therefore, the present invention provides a wind turbine power interval prediction method based on multi-layer decomposition and kernel density quantiles. The specific steps are as follows:

[0005] Step 1: Obtain a wind turbine power dataset and filter relevant feature data; preprocess the data and divide it into a training set and a test set, and construct an input feature vector based on the acquired data.

[0006] Step 2: Perform variational mode decomposition on the sequence and optimize the key parameters of VMD.

[0007] Step 3: Construct a PMGRU prediction model for wind turbine power interval prediction.

[0008] Step 4: Use the training set to train the PMGRU prediction model to obtain a trained model.

[0009] Step 5: Run the trained interval prediction model on the test set to obtain the prediction results.

[0010] Based on the above technical solution, the present invention produces the following beneficial effects:

[0011] Aiming at the problems of non-stationary information extraction, weakened time continuity and local extreme values ​​in the optimization process of existing wind power forecasting methods, the present invention uses variational modulus decomposition and kernel density loss to construct a wind turbine power interval prediction method based on SAPSO optimization, which realizes the adaptive extraction of sequence uncertainty and improves the accuracy and generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flowchart of a method for predicting wind turbine power intervals based on multi-level decomposition and kernel density quantiles according to an embodiment of the present application;

[0013] Figure 2 This is the basic framework and flow chart of the PMGRU prediction model in the embodiment of the present application;

[0014] Figure 3 It is the overall solution of the experimental example of this application. DETAILED DESCRIPTION

[0015] The specific embodiments of the present invention are described below in conjunction with the accompanying drawings to more clearly illustrate the technical solutions of the present invention.

[0016] like Figure 1-Figure 2 As shown, this embodiment provides a method for predicting wind turbine power intervals based on multi-level decomposition and kernel density quantiles based on deep learning. First, the principal component analysis method is used to analyze and filter the collected feature data, and the corresponding input matrix is ​​constructed based on fixed batches; secondly, the input matrix is ​​decomposed into subsequences of different frequencies through variational mode decomposition; thirdly, the PMGRU prediction model is trained to obtain the output result; finally, the prediction result is compared with the true value, and the evaluation indicators such as PICP and MPIW are used to evaluate the result. Specifically:

[0017] Step 1: Collect the characteristic data of the required data at each moment, and use the principal component analysis method to filter out some features with high correlation, such as wind speed, wind direction, temperature, air pressure and other environmental factors.

[0018] Preprocess the data, filter out missing values, fill them with median interpolation, and divide them into training set and test set; encode non-numeric data such as time; construct input feature vector based on the collected data, the feature vector X at time t t It can be expressed as:

[0019] x in={D t ,P t ,...,W t}

[0020] Among them D t represents the time information at time t, P t Indicates the output power information at time t, W t represents the wind speed information at time t, then the feature data set can be expressed as X in ={x1,x2,...,x t}∈R L×C , C represents the number of input features.

[0021] Step 2: Consider the sequence variables as time domain information and decompose them by variational mode decomposition (VMD). Before decomposition, the key parameters of VMD, such as frequency center, penalty factor, bandwidth, and number of modes, are selected by SAPSO optimization algorithm, and the sequence is decomposed to obtain subsequences VMF1, ..., VMF k When the particles are updated, the simulated annealing algorithm is introduced into the PSO algorithm to help PSO adaptively accept the optimal solution and avoid the gradient vanishing problem. The details are as follows:

[0022]

[0023] Where K is the modal number, u k (t) is the modal component, ω k is the center frequency.

[0024] Step 3: Construct a PMGRU prediction model for wind turbine power interval prediction.

[0025] First, when constructing the model input, an input matrix is ​​constructed according to the batch size, and the sequence is decomposed through the MovAvg module to obtain seasonal and trend components. The trend component is input into the gated recurrent network to capture the temporal dependency of the sequence, and the sparse attention mechanism is used to focus on the global sequence and screen the top k attention weight coefficients. A MoveAvg operation is performed after each attention module calculation is completed. Finally, the sequence is output as a prediction interval through a fully connected layer.

[0026] Furthermore, the PMGRU prediction model includes five key layers: a sparse attention module, a GRU gating network, a MovingAvg convolutional smoothing module, a residual connection module, and a fully connected layer. Key parameters in the PMGRU prediction model include convolution kernel size, dropout, initial learning rate, batch size, number of hidden layers, and loss function.

[0027] Among them, the sparse attention module is used to screen the attention weight coefficient; the GRU gating network is used to capture the temporal dependency of the sequence; the MovingAvg module is used to gradually extract the long-term stable trend from the sequence; the residual connection module is used to combine the fluctuation information to capture complex temporal patterns; the fully connected layer is used to project the sequence into the prediction interval.

[0028] Based on this, Figure 2 As shown, the working method of the PMGRU prediction model is:

[0029] Input the sequence into the ReVIN module and reorder the sequence to facilitate subsequent operations. At this time, the sequence dimension is The GRU gating network is used to extract temporal dependencies between features. Seasonal decomposition (MovingAvg module) is introduced as an internal operation of the PMGRU prediction model, which can gradually extract long-term stable trends from the sequence. For the sequence X, it can be expressed as:

[0030] X l =AvgPool(Padding(X))

[0031] trend=XX l

[0032] Among them, X l This example uses AvgPool(·) to extract this mild trend and maintains the same length as the original sequence by padding.

[0033] In order to reduce the influence of non-stationary factors when processing the coupling relationship between variables, this embodiment adopts a self-attention mechanism based on a sparse method. Figure 2 As shown in Figure 2, unlike the traditional attention mechanism, sparse attention can select the top k weights, reducing the computational complexity without affecting the accuracy. The attention is calculated as follows:

[0034] Q,K,V=Projection(x)

[0035] A(Q,K,V)=softmax(Scores[:k])·V

[0036] Among them, Query (Q), Key (K), and Values ​​(V) are transformed by the projection(·) function; Scores[:k] represents the selection of the first k attention weights, and the attention weight calculation can be expressed as A represents the attention weight.

[0037] To help the GRU network effectively identify fluctuation information and stable trends in the sequence, the PMGRU structure combines the AvgPool(·) fluctuation information through the residual connection of each layer of the GRM network to capture complex temporal patterns. The mathematical expression is as follows:

[0038] X l =LayerNorm(GRU(X l-1 )+trend l-1 )

[0039] Where: GRU(·) represents the calculation process of each block in GRU, trend represents the result of MovingAvg, and LayerNorm(·) represents the normalization operation of the layer.

[0040] Finally, the sequence X is transformed into l Projected into the prediction interval Y.

[0041] Step 4: Use the historical data of wind turbine power to train the PMGRU prediction model to obtain a trained model.

[0042] This example uses the Adam optimizer, using the training set mentioned in step S1, to continuously update the learning rate during training to improve the model's prediction accuracy. The model employs a quantile-based loss function, incorporating probability density information into the loss function to quantify the prediction error. Furthermore, the KDE weights are dynamically adjusted based on the density of the target value.

[0043] In particular, this embodiment constructs a kernel density-based quantile interval loss function during training, introducing probability density information into the loss function to quantify prediction errors. Kernel density estimation (KDE) is used to estimate the probability density distribution of the predicted values, enabling the model to consider distributional characteristics and construct different levels of attention. In addition, KDE weights are dynamically adjusted based on the density of the target value, making the model more adaptable to the actual data distribution, thereby improving the model's stability and accuracy.

[0044]

[0045] Where n is the number of samples, is the KDE function, K is the kernel function, and h is the bandwidth parameter.

[0046] Based on KDE adaptive adjustment, the complexity of the regularization control model is introduced to adjust the interval width. The final loss function can be expressed as:

[0047]

[0048] Among them, △ i It represents the width of the prediction interval of step i, which can be expressed as λ is a hyperparameter that controls the strength of regularization, and N represents the length.

[0049] Step 5: Run the trained interval prediction model on the test set to obtain the prediction results. The prediction results are output in the form of [L, U], where L represents the lower limit of the interval and U represents the upper limit of the interval.

[0050] Step 6: Use the test set to evaluate model performance, calculate the prediction error, and conduct an overall assessment of the model by analyzing the coverage ratio of the true value to the prediction interval and the average interval width. Use metrics such as prediction interval coverage probability (PICP), mean prediction interval width (MPIW), and inference time to assess model accuracy.

[0051]

[0052] Among them, y i represents the true value, U i and L i represents the upper and lower limits of interval i; n is the number of samples, [L i ,U i ] is the prediction interval; I(·) represents the indicator function, which is 1 if y is within the interval and 0 otherwise; α is the confidence level.

[0053] like Figure 3 As shown, this application also provides an experimental example. The prediction scheme of this experimental example mainly includes three parts: data acquisition, data preprocessing and model construction.

[0054] The steps 1 and 2 are specifically as follows:

[0055] The relevant characteristic data of the wind turbines is collected at a fixed time frequency at each moment, and a characteristic dataset is constructed based on the time information. The dataset includes the date, wind turbine output power, status characteristics (temperature, blade speed, height), and weather characteristics (air pressure, wind speed, air density, wind direction, etc.).

[0056] The wind turbine output power at the current moment is used as the target variable, and the training set and test set are divided.

[0057] Perform principal component analysis (PCA) on the collected data to screen features with high correlation; perform anomaly detection on the screened feature data set, and perform median interpolation on missing values ​​and outliers; encode non-numerical features such as dates to facilitate model recognition and make full use of time series information.

[0058] Perform reversible instance normalization on the dataset:

[0059]

[0060] Among them, Xt is the time series data; μ is the mean; σ is the standard deviation. The denormalization operation is

[0061] Construct the corresponding feature vector X t :

[0062] X in ={x1,x2,...,x t}∈R L×C

[0063] X t ={D t ,P t ,...,W t}

[0064] Among them, D t represents the time information at time t, P t Indicates the output power information at time t, W t Represents the wind speed information at time t; L represents the sequence length, and C represents the number of input features.

[0065] from Figure 3 Step 3 shows the training process of the prediction model in detail. The relevant parameters of VMD are optimized through SAPSO to decompose the subsequences that reflect the data trend; the data is input into PMGRU, and the model is trained through the KDE-Pinball loss parameter to finally output the prediction interval.

[0066] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A wind turbine power interval prediction method based on multi-level decomposition and kernel density quantile, characterized in that: The following steps are involved: Step 1: Obtain a wind turbine power dataset and filter relevant feature data; preprocess the data and divide it into a training set and a test set, and construct an input feature vector based on the acquired data; Step 2: Perform variational mode decomposition on the sequence and optimize key parameters; Step 3: Construct a PMGRU prediction model for wind turbine power interval prediction; Step 4: Use the training set to train the PMGRU prediction model to obtain a trained model; Step 5: Run the trained interval prediction model on the test set to obtain the prediction results.

2. The wind turbine power interval prediction method based on multi-level decomposition and kernel density quantile according to claim 1 is characterized in that: The step 1 is specifically as follows: Obtain a wind turbine power dataset, use the current wind turbine output power as the target variable, and divide it into a training set and a test set; Use principal component analysis to screen relevant feature data; Perform anomaly detection on the filtered feature data set, perform median interpolation on missing values ​​and outliers, and encode non-numerical features; Perform reversible instance normalization on the dataset; Construct a feature vector based on the acquired data.

3. The wind turbine power interval prediction method based on multi-level decomposition and kernel density quantile according to claim 2 is characterized in that: The data set includes date, wind turbine output power, state characteristics and weather characteristics.

4. A wind turbine power interval prediction method based on multi-level decomposition and kernel density quantile according to claim 1 or 3, characterized in that: The step 2 is specifically as follows: The sequence variables are regarded as time domain information and variational mode decomposition is performed. Before decomposition, the key parameters of VMD are selected by the SAPSO optimization algorithm, and the sequence is decomposed to obtain subsequences. When updating particles, a simulated annealing algorithm is introduced to avoid the gradient vanishing problem.

5. The wind turbine power interval prediction method based on multi-level decomposition and kernel density quantile according to claim 1 is characterized in that: The PMGRU prediction model includes a sparse attention module, a GRU gating network, a MovingAvg module, a residual connection module and a fully connected layer; The sparse attention module is used to filter the attention weight coefficient; The GRU gating network is used to capture the temporal dependencies of the sequence; The MovingAvg module is used to gradually extract the long-term stationary trend from the sequence; The residual connection module is used to combine fluctuation information to capture complex temporal patterns; The fully connected layer is used to project the sequence into the prediction interval.

6. The wind turbine power interval prediction method based on multi-level decomposition and kernel density quantile according to claim 5 is characterized in that: The working method of the PMGRU prediction model is: Extract the temporal information of the sequence through the GRU gating network; The MovingAvg module uses a fixed convolution kernel to smoothly decompose the trend component and seasonal component. The trend component is input into a multi-layer sparse attention mechanism and a gating network to continuously focus on the global trend and extract time series information. The trend component is merged with the seasonal component and the prediction interval is output through a fully connected layer.

7. A wind turbine power interval prediction method based on multi-level decomposition and kernel density quantile according to claim 5 or 6, characterized in that: The key parameters in the PMGRU prediction model include convolution kernel size, Dropout, initial learning rate, batch size, number of hidden layers and loss function.

8. The wind turbine power interval prediction method based on multi-level decomposition and kernel density quantile according to claim 1 is characterized in that: In step 4, a quantile interval loss function based on kernel density is used in the training process, and probability density information is introduced into the loss function to quantify the prediction error.

9. The wind turbine power interval prediction method based on multi-level decomposition and kernel density quantile according to claim 1 is characterized in that: Also includes: Step 6: Use the test set to evaluate the prediction performance of the PMGRU prediction model, calculate the prediction error, and use the evaluation indicators to evaluate the model as a whole.

10. The wind turbine power interval prediction method based on multi-level decomposition and kernel density quantile according to claim 9 is characterized in that: The evaluation indicators include: prediction interval coverage probability PICP and average prediction interval width MPIW, and the calculation formula is: Among them, y i represents the true value, U i and L i represents the upper and lower limits of interval i; n is the number of samples, [L i ,U i ] is the prediction interval; I(·) represents the indicator function, which is 1 if y is within the interval and 0 otherwise; α is the confidence level.

Images