New energy station power interval prediction method

Abstract

The invention belongs to the technical field of new energy station power prediction, and particularly relates to a new energy station power interval prediction method. In order to overcome the defect that the prediction accuracy and precision of an existing new energy station power interval prediction method have the improvement space, the invention adopts the following technical scheme: the new energy station power interval prediction method comprises the following steps: constructing a VMD-SA-TCN-BiLSTM-KDE prediction model; historical output power data of a wind power plant station are obtained, the historical output power data are preprocessed and then input into SA-TCN-BiLSTM for training, and the preprocessing comprises the step of decomposing the data through VMD; the part of historical output power data obtained after preprocessing is input into the trained SA-TCN-BiLSTM, and a prediction result is obtained; and performing KDE kernel density estimation on the prediction error, and obtaining a historical output power prediction interval of the wind power plant station according to a prediction result and the prediction error. The method has the beneficial effect of higher prediction precision.

Description

Technical Field

[0001] This invention belongs to the field of power prediction technology for new energy power plants, and specifically relates to a method for predicting power ranges for new energy power plants. Background Technology

[0002] New energy power plants have become a core component of power supply. However, the output of new energy sources is affected by meteorological conditions such as wind speed and temperature, exhibiting highly intermittent and random characteristics. This uncertainty poses a severe challenge to the optimal dispatching of the power grid.

[0003] Power range prediction for renewable energy power plants involves forecasting the upper and lower boundaries of renewable energy output at a given confidence level. This is a crucial method for improving the economic efficiency of the power system and reducing grid security risks. However, renewable energy output is highly uncertain due to various factors, making it difficult for a single model to effectively capture the complex characteristics of renewable energy power output. Furthermore, noise interference in the raw data can lead to excessively wide prediction ranges or insufficient confidence levels.

[0004] To address this issue, prediction structures combining signal decomposition and deep neural networks have been widely adopted, but they still have the following drawbacks: 1. Empirical Mode Decomposition (EMD) suffers from mode aliasing, requiring manual selection of effective components and increasing computational burden; 2. Deep neural networks do not fully combine the advantages of various neural networks, relying on a single structure, making it difficult to simultaneously model local details and global trends, resulting in low efficiency in time series modeling; 3. They have weak characterization capabilities for complex time series features (such as the intermittency and abrupt changes in wind and solar power); 4. There is still room for improvement in the accuracy of power range prediction for new energy power plants.

[0005] Related literature: Patent application 1, CN108717579A - A method for predicting short-term wind power range; Patent application 2, CN118503724A - A method for predicting wind power and photovoltaic power range based on the combination of mode decomposition and hybrid neural network; Patent application 3, CN118508430A - A short-term wind power prediction method based on deep learning.

[0006] Related terms: Variational mode decomposition (VMD); Self-attention (SA) mechanism; Temporal Convolutional Network (TCN); Bidirectional Long Short-Term Memory (BiLSTM) network; Kernel density estimation (KDE). Summary of the Invention

[0007] This invention addresses the shortcomings of existing power range prediction methods for new energy power plants, which have room for improvement in prediction accuracy. It provides a power range prediction method for new energy power plants based on VMD-SA-TCN-BILSTM-KDE. First, the variational mode decomposition (WMD) method is used to process the raw output power data of the wind farm. Then, a bidirectional long short-term memory temporal convolutional neural network (TCN-BiLSTM) based on a self-attention mechanism (SA) is used to perform deterministic predictions on each frequency term. The kernel density estimation method (KDE) is then used to estimate the probability density of the prediction error of the deterministic prediction results. Finally, the prediction intervals are superimposed with the predicted values ​​to obtain the overall range prediction result. Compared with other classic range prediction models, this method achieves higher prediction accuracy.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: a power range prediction method for new energy power plants, the power range prediction method for new energy power plants comprising the following steps: S1. Construct the VMD-SA-TCN-BiLSTM-KDE prediction model; S2. Obtain historical output power data of wind farms and stations, preprocess it, and input it into SA-TCN-BiLSTM for training to obtain the trained SA-TCN-BiLSTM. The preprocessing includes using the variational mode decomposition method (VMD) to decompose the historical output power data of wind farms and stations. S3. Input the preprocessed historical output power data into the trained SA-TCN-BiLSTM model to obtain the prediction results; S4. Perform KDE kernel density estimation on the prediction error, and obtain the historical output power prediction range of the wind farm based on the prediction results and prediction error.

[0009] As an improvement, in S2, the time scale for historical output power data is a whole year; output power data is collected every 10 minutes.

[0010] As an improvement, in S2, the Variational Mode Decomposition (VMD) method is used to decompose the historical output power data of wind farms into three intrinsic mode functions (EMFs). To ensure that the decomposed sequence consists of mode functions with a finite bandwidth and a center frequency, and that the sum of the bandwidths of each mode function is minimized, the constraint condition is that the sum of all mode functions is equal to the power time series. The VMD constraint condition is that the sum of all mode functions is equal to the power time series. By continuously iterating and updating to find the optimal solution, the frequency of the output power data of wind farm stations can be extracted and separated to obtain the high-frequency, mid-frequency, low-frequency and residual terms of the corresponding output power data.

[0011] As an improvement, in S2, the preprocessing also includes standardizing the high-frequency, mid-frequency, low-frequency and residual terms of the output power data. After standardization, the dataset is scaled to a distribution range with a mean of 0 and a standard deviation of 1, so that the values ​​of each modal component are at similar orders of magnitude, unifying the scale of multimodal components and accelerating the convergence speed of the model.

[0012] As an improvement, in S2, each frequency term of the standardized wind farm power output data is used as the input data of SA-TCN-BiLSTM, and the output data of SA-TCN-BiLSTM is de-standardized to obtain the prediction result of the wind farm power output.

[0013] As an improvement to S2, the standardized wind farm power output data is divided into training set, validation set and test set according to a certain ratio. In S2, the training set data is input into SA-TCN-BiLSTM for training, and the weight parameters in the training model are minimized to minimize the mean square error between the predicted value and the true value. In S3, the validation set data is input into SA-TCN-BiLSTM-KDE to perform statistical analysis on the prediction error and find the confidence interval range; The power range prediction method for new energy power plants also includes: S5, which inputs the test set data into SA-TCN-BiLSTM-KDE to evaluate the model.

[0014] As an improvement, in S2, the self-attention mechanism SA dynamically assigns weights to the input time series output power data, captures the dependencies between different positions within the time series, and enhances attention to key parts. The Temporal Convolutional Network (TCN) employs a 4-layer dilated convolutional structure, with the dilation factor increasing by a factor of 1 / 2 in each layer, to extract long-term trend features from the time series. The Bidirectional Long Short-Term Memory (BiLSTM) network consists of two LSTMs, which process time-series data from the forward and backward directions respectively. It has 32 hidden units and a total of 64 units in the bidirectional structure, which fully learns the sequential dependencies of time series.

[0015] As an improvement, the probability density function used in S4 is expressed as follows:

[0016]

[0017] in, n To predict the number of samples in one attempt, h The bandwidth function is used to control the degree of smoothness. K For Gaussian kernel function, This represents the average prediction error. The actual value; This is a predicted value; By fitting the distribution of the prediction error, the probability density function of the prediction error is obtained.

[0018] As an improvement, in S4, the probability density function is further integrated to obtain the cumulative distribution function of the prediction error probability, as shown in the following expression: ; Based on the obtained cumulative distribution function of prediction error probability, the statistical characteristics of the error are quantified and confidence intervals are generated, as shown in the following expression:

[0019]

[0020] in, Let be the probability of the confidence interval. This represents the upper bound of the corresponding confidence interval. This represents the lower bound of the corresponding confidence interval. The predicted value is for deterministic forecasting; By statistically modeling the prediction error and combining it with the prediction results of the training set, the prediction range of the output power of the wind farm station is obtained.

[0021] As improvements, evaluation metrics include PICP, PINAW, CWC, and MPICD; The PICP metric is used to evaluate the coverage of probability interval prediction models. The PINAW metric is used to evaluate the average width of the prediction interval of a probability interval prediction model. By comparing it with the range of observed values, it can be used to judge the prediction accuracy and generalization ability of the model. The CWC indicator takes into account both PICP and PINAW, that is, it takes into account the coverage and narrowing of the prediction interval. The MPICD metric is used to evaluate the degree of deviation between the midpoint of the prediction interval and the actual value.

[0022] The new energy power station power range prediction method of the present invention processes the data sequentially using VMD, SA, TCN, BILSTM and KDE, and has the following beneficial effects: (1) The variational mode decomposition (VMD) adopted can adaptively decompose the power sequence of new energy power plants. By constraining the center frequency and bandwidth of the mode components, it overcomes the mode aliasing problem of traditional empirical mode decomposition and obtains clearer high and low frequency subsequences.

[0023] (2) Self-attention mechanism (SA) is used in time series prediction to enhance the dependency between different positions within the time series. By assigning different weights to each position, the model’s attention to key parts is enhanced.

[0024] (3) The Temporal Convolutional Network (TCN) used expands the receptive field through dilated causal convolution, and achieves efficient capture of long-distance dependencies while ensuring temporal causality. It utilizes the advantages of parallel computing to improve learning efficiency and has stronger temporal modeling efficiency than the local feature extraction of traditional CNN networks.

[0025] (4) The bidirectional long short-term memory network BiLSTM is used to capture forward and backward time-series dynamic characteristics simultaneously through bidirectional structure. Compared with the unidirectional LSTM network, it has significantly enhanced the ability to represent complex time-series characteristics (such as the intermittency and abrupt changes of wind and solar power). By integrating historical and future implicit state information, it can more accurately model the periodic fluctuations and long-term trend changes in power sequences.

[0026] (5) An adaptive bandwidth optimization algorithm is introduced into the probability prediction of the kernel density estimation KDE. By dynamically adjusting the kernel function bandwidth, the uncertainty of the new energy power range can be characterized more precisely. Attached Figure Description

[0027] Figure 1 This is a flowchart of the power range prediction method for new energy power plants according to an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the process of predicting the power range of new energy power plants according to an embodiment of the present invention.

[0029] Figure 3 This is a schematic diagram of the process of predicting the power range of new energy power plants according to an embodiment of the present invention (the internal structure of the model is also shown).

[0030] Figure 4 This is a comparison chart of the predicted interval and the actual value obtained using the method of Embodiment 1 of the present invention.

[0031] Figure 5 This is a comparison chart of the predicted interval and the true value obtained using the SA-CNN-LSTM-KDE model.

[0032] Figure 6 This is a comparison chart of the predicted interval and the true value obtained using the VMD-SA-CNN-LSTM-KDE model. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be explained and described below. However, the following embodiments are only preferred embodiments of the present invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of the present invention.

[0034] See Figure 1 The power range prediction method for new energy power plants according to Embodiment 1 of the present invention includes the following steps: S1. Construct the VMD-SA-TCN-BiLSTM-KDE prediction model; S2. Obtain historical output power data of wind farms and stations, preprocess it, and input it into SA-TCN-BiLSTM for training to obtain the trained SA-TCN-BiLSTM. The preprocessing includes using the variational mode decomposition method (VMD) to decompose the historical output power data of wind farms and stations. S3. Input the preprocessed historical output power data into the trained SA-TCN-BiLSTM model to obtain the prediction results; S4. Perform KDE kernel density estimation on the prediction error, and obtain the historical output power prediction range of the wind farm based on the prediction results and prediction error.

[0035] See Figure 2 The main idea of ​​the prediction method in this embodiment of the invention is as follows: First, obtain the historical output power data of a single wind turbine, and then summarize the historical output power data of the wind farm. Then, VMD is used to process the historical output power data of the wind farm to obtain high-frequency, mid-frequency, low-frequency and residual terms; Secondly, the data processed by VMD is input into the SA-TCN-BiLSTM model for prediction to obtain deterministic results (point values). Finally, the interval prediction (interval value) is obtained and output based on the KDE kernel density interval prediction and deterministic results.

[0036] See Figure 3 Historical output power data for wind farms is obtained by collecting data from individual wind turbines. Specifically, this can be achieved by collecting the output power data of a single wind turbine every 10 minutes.

[0037] Then, the historical output power data of the wind farm was divided into training, validation, and test sets in a 5:2:3 ratio. The training set data was used to train the weight parameters in the neural network to minimize the mean squared error between the predicted and actual values; the validation set was used for statistical analysis of the prediction error to find the confidence interval range; and the test set was used to evaluate the model's generalization performance based on PICP, PINAW, CWC, and MPICD metrics.

[0038] Simultaneously, based on the wind farm's output power data, VMD is used to decompose the historical output power data of the wind farm into three intrinsic mode functions. To ensure that the decomposed sequence consists of mode functions with a finite bandwidth and a center frequency, and that the sum of the bandwidths of each mode function is minimized, the constraint condition is that the sum of all mode functions is equal to the power time series. The VMD constrained variational model is as follows:

[0039] in, For modal functions; The center frequency of the modal function; For Dirac functions; This is a dataset of historical output power from wind farms.

[0040] Find the optimal solution through continuous iteration and updates. , This allows for the extraction and separation of the frequency of the output power data from wind farms, obtaining the high-frequency (IMF1), mid-frequency (IMF2), low-frequency (IMF3), and residual (res) terms of the corresponding output power data.

[0041] Furthermore, the high-frequency, mid-frequency, low-frequency, and residual terms of the output power data are standardized using the following formula:

[0042]

[0043]

[0044] in, The input time series dataset; The mean of the dataset; The standard deviation of the dataset; This is a standardized time series dataset.

[0045] After standardization, the dataset is scaled to a distribution range with a mean of 0 and a standard deviation of 1, so that the values ​​of each modal component are at similar magnitudes, unifying the scale of multimodal components and accelerating the convergence speed of the model.

[0046] Then, the high-frequency, mid-frequency, low-frequency, and residual data of the standardized wind farm power output are input into the SA-TCN-BILSTM for training. The SA-TCN-BILSTM model is trained based on the training set data. The time series input to the SA-TCN-BILSTM consists of the frequency terms of the standardized wind farm power output data (training set), and the output data, after destandardization, becomes the predicted wind farm power output for one day later.

[0047] Temporal Convolutional Neural Networks (TCNs) capture the long-term trend and local fluctuation features of the output power sequence through convolution. Dilated convolution is used to increase the coverage of the receptive field. In addition, residual connections are used to alleviate the gradient vanishing problem in deep networks. The expression for dilated convolution is shown below:

[0048] in, For the first in the output sequence One element, For the first in the input sequence One element, As the expansion factor, The location of the convolution kernel. N The kernel size is [size]. These are the kernel weights.

[0049] The residual block consists of dilated convolution, weight normalization, activation function, and dropout layer. It is directly added to the input through skip connections. Weight normalization accelerates convergence and improves model stability. The dropout layer randomly discards some neurons with a probability of 0.2 to prevent overfitting.

[0050] Bidirectional Long Short-Term Memory (BiLSTM) networks are used to capture the temporal relationships of power sequences, processing time-series data from both forward and backward directions. The internal structure of an LSTM unit includes an input gate, a forget gate, and an output gate. These mechanisms allow the LSTM unit to retain learned information for extended periods, better learning long-term dependencies in time series data while avoiding the vanishing and exploding gradient problems. Its expression is shown below:

[0051]

[0052]

[0053]

[0054]

[0055] in, , , These are the weight matrices for the input gate, forget gate, and output gate, respectively. , , These are the output sequences of the input gate, forget gate, and output gate, respectively. , , For the corresponding bias; Candidate cell states represent new information that can be added to the cell state; This represents the cell state at the current time step, containing information processed by the forget gate and the input gate. Input the sequence for the current time step; This is the previous time step hidden state; is the sigmoid activation function; tanh is the hyperbolic tangent activation function.

[0056] Self-attention (SA) is used in time series prediction to enhance the dependencies between different positions within a time series. By assigning different weights to each position, it enhances the model's attention to key parts. In SA, the input sequence is mapped to three different spatial vectors. The keys measure the relevance of the query to other elements in the sequence; the query identifies the part of the sequence of interest; and the values ​​are the output features obtained by weighted summation of the attention weights. The calculation expression is as follows:

[0057]

[0058]

[0059]

[0060] in, , , These are the query matrix, key matrix, and value matrix, respectively. , , This is the corresponding weight matrix; Input time series; This is the scaling factor; Score for attention; These are the weighting coefficients; This is the final attention output. SA, TCN, and BILSTM can all use existing models.

[0061] After training, the validation set data is input into the trained SA-TCN-BILSTM to obtain prediction results. Then, KDE kernel density estimation is performed on the prediction error. The probability density function used is shown below:

[0062]

[0063] in, n To predict the number of samples in one attempt, h The bandwidth function is used to control the degree of smoothness. K For Gaussian kernel function, This represents the average prediction error. The actual value; These are predicted values.

[0064] By fitting the prediction error distribution, the probability density function of the prediction error can be obtained. Further integration of the probability density function yields the cumulative probability distribution function of the prediction error, as shown in the following expression:

[0065] Based on the obtained cumulative distribution function of prediction error probability, the statistical characteristics of the error are quantified and confidence intervals are generated, as shown in the following expression:

[0066]

[0067] in, The probability of a confidence interval. This represents the upper bound of the corresponding confidence interval. This represents the lower bound of the corresponding confidence interval. The predicted value is for deterministic prediction.

[0068] S3. After adjusting the parameters of SA-CNN-SLTM using the validation set, input the test set into SA-TCN-BiLSTM-KDE to obtain the prediction results, and use PICP, PINAW, CWC, and MPICD metrics to measure and evaluate the prediction results.

[0069]

[0070] in I This is an indicator function; it takes the value 1 when the condition within the parentheses is true, and 0 otherwise. This formula represents the probability that the actual value falls within the prediction interval. A higher PICP indicates a more reliable prediction interval and better predictive ability.

[0071]

[0072] By calculating PINAW, we can evaluate the proportion of the average width of the probability interval prediction model to the true value range, and thus judge the prediction accuracy and generalization ability of the model. The smaller the value, the smaller the proportion of the prediction interval to the true value range, and the higher the prediction accuracy.

[0073]

[0074] Because there is a certain contradiction between the PICP and PINAW indices—a larger PINAW value tends to correlate with a larger PICP value—CWC comprehensively considers both the coverage and narrowing rate of the prediction interval. This refers to the set confidence level, a 90% confidence level. =0.1; if PICP ≥ ,but The value is 0 if it is greater than the confidence interval, and 1 otherwise; when PICP is less than the confidence interval, If the value is 1, there exists a penalty coefficient. The penalty for prediction errors. Therefore, the smaller the CWC, the better the performance of the prediction model.

[0075]

[0076] This formula indicates that the closer the midpoint of the prediction interval is to the true value, the better the MPICD index is, and the more accurate the time series prediction is.

[0077] See Figures 4 to 6 ,Will Figure 4 The prediction interval obtained by the VMD-SA-TCN-BiLSTM-KDE model in this embodiment is similar to that of existing models (such as...). Figure 5 SA-CNN-SLTM-KDE, Figure 6 A comparison was made between VMD-SA-CNN-LSTM-KDE. Figures 4 to 6 As can be seen, the power prediction range obtained by VMD-SA-TCN-BiLSTM-KDE in this embodiment is closer to the actual value, verifying the effectiveness of the proposed power range prediction model VMD-SA-TCN-BILSTM-KDE. Compared with existing models, it can provide a more reliable and narrower confidence interval for power range prediction. Specifically, the model in this embodiment and existing models are compared using evaluation metrics (PICP, PINAW, and CWC), and the comparison results are as follows.

[0078] Table 1. Comparison of evaluation metrics for the 30% confidence interval of the prediction models

[0079] Table 2 Comparison of evaluation metrics for the 60% confidence interval of the prediction models

[0080] Table 3 Comparison of evaluation metrics for the 90% confidence interval of the prediction models

[0081] As can be seen from Tables 1 to 3, the PICP value of the VMD-SA-TCN-BILSTM-KDE model in this embodiment is higher than that of the existing VMD-SA-CNN-LSTM-KDE or SA-CNN-SLTM-KDE, indicating that the prediction range of the model in this embodiment is more reliable and the prediction ability is better; PINAW is not much different from the existing models; CWC is significantly smaller than the existing models, indicating that the model has better performance.

[0082] The power range prediction method for new energy power plants in Embodiment 1 of this invention constructs a prediction model based on VMD-SA-TCN-BiLSTM-KDE. First, VMD decomposes the power sequence of new energy power plants into multiple subsequences with different frequency characteristics, effectively decomposing the nonlinear components in the data and providing time-domain and frequency-domain information references for refined modeling of new energy power plant power. Addressing the long-term characteristics of new energy power plant power driven by seasonal changes in sunlight and wind speed, TCN is used to stack multiple convolutional layers with increasing expansion coefficients, achieving efficient capture of multi-scale global features of the input time series. To further characterize the short-term intermittent fluctuations in new energy power plant power, BiLSTM is introduced based on the high-dimensional global features extracted by TCN. A bidirectional gating mechanism combines forward and backward time dimension information, enhancing the model's learning ability for local features. Furthermore, an attention mechanism SA is introduced into the neural network to capture the dependencies between different positions within the power sequence of new energy power plants. By assigning different weights to each position, the model's attention to key parts is enhanced, resulting in better learning ability for the complementary features exhibited by the power sequence of new energy power plants in space and time.

[0083] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art should understand that the present invention includes, but is not limited to, the content described in the above specific embodiments. Any modifications that do not depart from the functional and structural principles of the present invention will be included within the scope of the claims.

Claims

1. A method for predicting the power range of new energy power plants, characterized in that: The method for predicting the power range of new energy power plants includes the following steps: S1. Construct the VMD-SA-TCN-BiLSTM-KDE prediction model; S2. Obtain historical output power data of wind farms and stations, preprocess it, and input it into SA-TCN-BiLSTM for training to obtain the trained SA-TCN-BiLSTM. The preprocessing includes using the variational mode decomposition method (VMD) to decompose the historical output power data of wind farms and stations. S3. Input the preprocessed historical output power data into the trained SA-TCN-BiLSTM model to obtain the prediction results; S4. Perform KDE kernel density estimation on the prediction error, and obtain the historical output power prediction range of the wind farm based on the prediction results and prediction error.

2. The method for predicting the power range of new energy power plants according to claim 1, characterized in that: In S2, the time scale for historical output power data is one full year; output power data is collected every 10 minutes.

3. The method for predicting the power range of new energy power plants according to claim 1, characterized in that: In S2, the variational mode decomposition (VMD) method is used to decompose the historical output power data of the wind farm into three intrinsic mode functions (EMFs). To ensure that the decomposed sequence consists of mode functions with a finite bandwidth and a center frequency, and that the sum of the bandwidths of each mode function is minimized, the constraint condition is that the sum of all mode functions is equal to the power time series. The VMD constraint condition is that the sum of all mode functions is equal to the power time series. By continuously iterating and updating to find the optimal solution, the frequency of the output power data of wind farm stations can be extracted and separated to obtain the high-frequency, mid-frequency, low-frequency and residual terms of the corresponding output power data.

4. The method for predicting the power range of new energy power plants according to claim 3, characterized in that: In S2, preprocessing also includes standardizing the high-frequency, mid-frequency, low-frequency and residual terms of the output power data. After standardization, the dataset is scaled to a distribution range with a mean of 0 and a standard deviation of 1, so that the values ​​of each modal component are at similar orders of magnitude, unifying the scale of multimodal components and accelerating the convergence speed of the model.

5. The method for predicting the power range of new energy power plants according to claim 4, characterized in that: In S2, each frequency term of the standardized wind farm power output data is used as the input data of SA-TCN-BiLSTM. The output data of SA-TCN-BiLSTM is de-standardized to obtain the prediction result of the wind farm power output.

6. The method for predicting the power range of new energy power plants according to claim 5, characterized in that: In S2, the standardized wind farm power output data is divided into training set, validation set and test set according to a certain ratio. In S2, the training set data is input into SA-TCN-BiLSTM for training, and the weight parameters in the training model are minimized to minimize the mean square error between the predicted value and the true value. In S3, the validation set data is input into SA-TCN-BiLSTM-KDE to perform statistical analysis on the prediction error and find the confidence interval range; The power range prediction method for new energy power plants also includes: S5, which inputs the test set data into SA-TCN-BiLSTM-KDE to evaluate the model.

7. The method for predicting the power range of new energy power plants according to claim 6, characterized in that: In S2, the self-attention mechanism SA dynamically assigns weights to the input time series output power data, captures the dependencies between different positions within the time series, and enhances attention to key parts. The Temporal Convolutional Network (TCN) employs a 4-layer dilated convolutional structure, with the dilation factor increasing by a factor of 1 / 2 in each layer, to extract long-term trend features from the time series. The Bidirectional Long Short-Term Memory (BiLSTM) network consists of two LSTMs, which process time-series data from the forward and backward directions respectively. It has 32 hidden units and a total of 64 units in the bidirectional structure, which fully learns the sequential dependencies of time series.

8. The method for predicting the power range of new energy power plants according to claim 6, characterized in that: In S4, the probability density function used is expressed as: in, n To predict the number of samples in one attempt, h The bandwidth function is used to control the degree of smoothness. K For Gaussian kernel function, This represents the average prediction error. The actual value; This is a predicted value; By fitting the distribution of the prediction error, the probability density function of the prediction error is obtained.

9. The method for predicting the power range of new energy power plants according to claim 8, characterized in that: In S4, the probability density function is further integrated to obtain the cumulative distribution function of the prediction error probability, as shown in the following expression: ； Based on the obtained cumulative distribution function of prediction error probability, the statistical characteristics of the error are quantified and confidence intervals are generated, as shown in the following expression: in, Let be the probability of the confidence interval. This represents the upper bound of the corresponding confidence interval. This represents the lower bound of the corresponding confidence interval. The predicted value is for deterministic forecasting; By statistically modeling the prediction error and combining it with the prediction results of the training set, the prediction range of the output power of the wind farm station is obtained.

10. The method for predicting the power range of new energy power plants according to claim 6, characterized in that: Evaluation metrics include PICP, PINAW, CWC, and MPICD; The PICP metric is used to evaluate the coverage of probability interval prediction models. The PINAW metric is used to evaluate the average width of the prediction interval of a probability interval prediction model. By comparing it with the range of observed values, it can be used to judge the prediction accuracy and generalization ability of the model. The CWC indicator takes into account both PICP and PINAW, that is, it takes into account the coverage and narrowing of the prediction interval. The MPICD metric is used to evaluate the degree of deviation between the midpoint of the prediction interval and the actual value.

Images