A plateau photovoltaic short-term point-interval prediction method, system, device and medium

Abstract

The application discloses a kind of plateau photovoltaic short-term point-interval prediction method, system, equipment and medium, it is related to photovoltaic power generation power prediction technical field, including obtaining plateau photovoltaic historical data to construct sample set, train point prediction model to obtain point prediction result;According to point prediction result and true value, construct error sequence, using rolling window slice and calculate the fuzzy distribution entropy of each window error;According to fuzzy distribution entropy dynamically adjusts reference bandwidth to construct adaptive bandwidth function, carries out kernel density estimation based on adaptive bandwidth function and calculates error quantile interval under pre-setting reliability;Error quantile interval is superimposed to point prediction result and is corrected according to physical output boundary, and the output prediction interval is output.The application introduces dynamic time warping distance in point prediction stage to improve loss function to optimize morphological fitting, and combines multi-head attention mechanism to enhance key feature capture ability, improves point prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, and in particular to a short-term point-interval prediction method for high-altitude photovoltaic power generation. Background Technology

[0002] Plateau regions, rich in photovoltaic resources, possess significant development potential. However, due to the influence of topography and complex weather conditions, photovoltaic power generation in these areas often exhibits strong randomness, volatility, and instability. This not only increases the difficulty of short-term power forecasting but also places higher demands on the optimized operation of the power system and the efficient absorption of new energy sources.

[0003] On the one hand, existing point prediction models have limited ability to characterize abrupt changes and key fluctuations in photovoltaic power, resulting in low accuracy in fitting the morphological features of the target daily sequence and in depicting the details of abrupt changes. Therefore, it is urgent to improve the traditional prediction model architecture by introducing a multi-head attention mechanism layer (MHA) and dynamic time warping distance (DTW) to enhance its short-term point prediction performance in response to the strong volatility of photovoltaic power in high-altitude areas. On the other hand, the prediction information provided by a single point prediction result is limited, only reflecting the fluctuation of photovoltaic power in the short term, and cannot provide comprehensive information for grid dispatch, reserve configuration, and other aspects. Therefore, it is necessary to conduct research on short-term interval prediction methods for photovoltaic power in high-altitude areas, using fuzzy distribution entropy (FDE) to achieve adaptive updates of the bandwidth parameters of traditional kernel density estimation (KDE), effectively improving the model's ability to characterize the error distribution when dealing with sequences of different complexities and volatility. Summary of the Invention

[0004] To address the shortcomings of the aforementioned technologies, this invention provides a short-term point-interval prediction method for high-altitude photovoltaic systems.

[0005] To solve the above technical problems, the technical solution adopted by this invention is: a short-term point-interval prediction method for high-altitude photovoltaic systems, comprising: Historical operational data of photovoltaic power generation on plateaus were obtained to construct a historical daily feature sample set.

[0006] A point prediction model is constructed, and the model is trained using a historical daily feature sample set. The point prediction results are obtained through the point prediction model.

[0007] Based on the point prediction results and the true values, a prediction error sequence is constructed, and a rolling time window is used to locally slice the error sequence to calculate the fuzzy distribution entropy of the error within each window.

[0008] Based on the fuzzy distribution entropy of each window, the baseline bandwidth is dynamically adjusted. An adaptive bandwidth function is constructed based on the adjusted baseline bandwidth. The kernel density of the error samples in each window is estimated based on the adaptive bandwidth function, and the error quantile interval is calculated under a preset confidence level.

[0009] The error quantile intervals are superimposed onto the point prediction results, and upper and lower limits are corrected according to the physical output boundary of photovoltaic power to output the short-term photovoltaic power prediction range.

[0010] Furthermore, a point prediction model is constructed and trained using a historical daily feature sample set, including: A feature extraction network is used to extract sequence features from a historical daily feature sample set, and the extracted feature information sequence is input into a time-series learning network to obtain a prediction vector with time-series features.

[0011] The predicted vector is input into the multi-head attention mechanism layer for weighted processing to obtain the feature vector after attention reconstruction.

[0012] The feature vector is input into the fully connected layer, and the dynamic time warp distance is used as a constraint term in the loss function to participate in the parameter update. After training, the point prediction result is output through the point prediction model.

[0013] Furthermore, the predicted vectors are input into the multi-head attention mechanism layer and weighted, including: The prediction vector is mapped to a query vector, a key vector, and a value vector.

[0014] Calculate the similarity between the query vector and the key vector, and normalize the similarity to obtain the attention weight.

[0015] The attention output is obtained by weighting and summing the value vectors using attention weights.

[0016] Parallel attention computation is performed in a parallel subspace using multiple attention heads, and the outputs of each attention head are concatenated and fused to obtain the feature vector after attention reconstruction.

[0017] Furthermore, the feature vectors are input into the fully connected layer, and the dynamic time-warped distance is used as a constraint term in the loss function to participate in parameter updates, including: The attention-reconstructed feature vectors are input into the fully connected layer to construct the total loss function. With the goal of minimizing the total loss function, the network parameters of the point prediction model are updated through backpropagation.

[0018] After training, the feature data of the target day is input into the point prediction model, and the point prediction results are output after forward calculation.

[0019] Furthermore, a prediction error sequence is constructed based on the point prediction results and the true values. A rolling time window is used to locally slice the error sequence, and the fuzzy distribution entropy of the error within each window is calculated, including:

[0020] Extract the predicted values ​​and corresponding true values ​​at each time point, and construct an absolute error sample by subtracting the values ​​at each point.

[0021] A rolling error window is set on the absolute error samples and slides sequentially along the time axis to divide the full-time error sequence into several local error subsequences.

[0022] Calculate the fuzzy distribution entropy of the error subsequence within each rolling window.

[0023] This preferred scheme extracts the fuzzy distribution entropy of each window through a rolling window, quantifies the time-varying complexity characteristics of the error sequence, and provides a refined measurement basis for bandwidth adaptive adjustment.

[0024] Furthermore, based on the fuzzy distribution entropy of each window, the baseline bandwidth is dynamically adjusted, and an adaptive bandwidth function is constructed based on the adjusted baseline bandwidth, including:

[0025] The reference bandwidth is determined based on the statistical characteristics of the error samples.

[0026] The fuzzy distribution entropy of each window is compared with the fuzzy distribution entropy of the global error sequence, and the adaptive bandwidth is obtained by adjusting the intensity adjustment coefficient.

[0027] Apply broadband boundary constraints to the adjusted adaptive bandwidth to keep the bandwidth value within a reasonable range.

[0028] This preferred solution achieves adaptive adjustment of bandwidth with local error complexity by comparing the window fuzzy distribution entropy with the global fuzzy distribution entropy and supplementing it with boundary constraints, thus improving the problem of poor applicability of fixed bandwidth.

[0029] Furthermore, kernel density estimation is performed on error samples within each window based on an adaptive bandwidth function, and error quantile intervals are calculated under a preset confidence level, including:

[0030] Based on adaptive bandwidth, and combined with error samples within the rolling window, the density function is obtained.

[0031] Integrating the density function yields the cumulative distribution function.

[0032] At a preset confidence level, the upper and lower quantiles corresponding to the confidence level are determined based on the cumulative distribution function.

[0033] The error interval is obtained based on the upper and lower quantiles.

[0034] This preferred scheme estimates kernel density based on adaptive bandwidth and solves for quantiles under a preset confidence level, thereby accurately characterizing the error distribution and improving the reliability and accuracy of interval prediction.

[0035] This invention provides a short-term point-to-interval prediction system for high-altitude photovoltaic systems.

[0036] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a plateau photovoltaic short-term point-interval prediction system, comprising: a data acquisition and preprocessing module, a point prediction model construction and training module, an error sequence construction module, an adaptive bandwidth adjustment module, a kernel density estimation and quantile interval calculation module, and an interval synthesis and boundary correction module.

[0037] The data acquisition and preprocessing module acquires historical operational data of plateau photovoltaic power generation and constructs a historical daily feature sample set.

[0038] The point prediction model construction and training module constructs a point prediction model, trains the point prediction model using a historical daily feature sample set, and obtains the point prediction results through the point prediction model.

[0039] The error sequence construction module constructs a prediction error sequence based on the point prediction results and the true values, and uses a rolling time window to locally slice the error sequence and calculate the fuzzy distribution entropy of the error within each window.

[0040] The adaptive bandwidth adjustment module dynamically adjusts the baseline bandwidth based on the fuzzy distribution entropy of each window, and constructs an adaptive bandwidth function based on the adjusted baseline bandwidth.

[0041] The kernel density estimation and quantile interval calculation module performs kernel density estimation on error samples within each window based on an adaptive bandwidth function, and calculates the error quantile interval under a preset confidence level.

[0042] The interval synthesis and boundary correction module superimposes the error quantile intervals onto the point prediction results and corrects them according to the physical output boundary of photovoltaic power, outputting the short-term photovoltaic power prediction interval.

[0043] Compared with the prior art, the beneficial effects of the present invention are: the present invention proposes a CNN-LSTM-MHA-DTW plateau photovoltaic short-term point prediction method based on dynamic time warping distance (DTW) and multi-head attention mechanism (MHA).

[0044] (1) The DTW distance index is added to the loss function of the prediction model to achieve dynamic correction of the morphological features of the predicted sequence. At the same time, the multi-head attention mechanism is used to collect key moments and features in parallel, which can effectively improve the model's capture effect on key mutation features and improve the accuracy of the model in the short-term point prediction scenario of plateau photovoltaic.

[0045] (2) This invention uses fuzzy distribution entropy (FDE) to measure the local complexity of the error sequence and dynamically corrects the kernel density estimation bandwidth based on this. This can effectively improve the problem that fixed bandwidth kernel density estimation is difficult to adapt to the differences in error distribution during different fluctuation periods, and can effectively improve the prediction effect of the model in the short-term interval prediction scenario of plateau photovoltaic. Attached Figure Description

[0046] Figure 1 This is an overall flowchart of the plateau photovoltaic short-term point-interval prediction method in this invention.

[0047] Figure 2 This is a flowchart of the multi-head attention mechanism in this invention.

[0048] Figure 3 This is a flowchart of the dynamic time warping distance in this invention. Detailed Implementation

[0049] To make the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0050] Example 1, referring to Figure 1 and Figure 2 This is one embodiment of the present invention, which provides a short-term point-interval prediction method for high-altitude photovoltaic systems, comprising: S1. Obtain historical operating data of plateau photovoltaic power generation and construct a historical daily feature sample set.

[0051] We collected a small sample of historical data on high-altitude photovoltaic power, including six meteorological data points: irradiance, temperature, humidity, wind speed, air pressure, and wind direction, as well as corresponding photovoltaic power data. We constructed a historical daily feature matrix of multiple meteorological elements throughout the year, which served as training data for the point prediction model.

[0052] S2. Construct a point prediction model, train the point prediction model using a historical daily feature sample set, and obtain the point prediction results through the point prediction model.

[0053] A convolutional neural network is used to extract sequential features from a historical daily feature sample set. The extracted feature information sequence is then input into a long short-term memory network to model its temporal dependence. The output prediction vector is input into a multi-head attention mechanism layer for weighted processing to strengthen the representation of key information. The attention-reconstructed feature vector is then input into a fully connected layer, and dynamic time warping distance is used as one of the constraints of the loss function to participate in the model parameter update. After training, a point prediction model is used to make point predictions for the target day, and the photovoltaic power point prediction results are output.

[0054] S3. Construct a prediction error sequence based on the point prediction results and the true values, and use a rolling time window to locally slice the error sequence, and calculate the fuzzy distribution entropy (FDE) of the error within each window.

[0055] Extract the point prediction results and corresponding true values ​​at each time point, construct absolute error samples by subtracting each point, set a rolling error window on the absolute error samples, slide it sequentially along the time axis, divide the full time domain error sequence into several local error subsequences, calculate the fuzzy distribution entropy of the error subsequence in each rolling window, and use the fuzzy distribution entropy to measure the complexity characteristics of the error sequence in each local time period.

[0056] S4. Based on the fuzzy distribution entropy of each window, the baseline bandwidth is dynamically adjusted. An adaptive bandwidth function is constructed based on the adjusted baseline bandwidth. Kernel density estimation (KDE) is performed on the error samples in each window based on the adaptive bandwidth function, and the error quantile interval is calculated under the preset confidence level.

[0057] The baseline bandwidth is determined based on the statistical characteristics of all error samples. The fuzzy distribution entropy of each window is compared with the fuzzy distribution entropy of the global error sequence and adjusted by the intensity adjustment coefficient. Bandwidth boundary constraints are applied to the adjusted adaptive bandwidth to keep the bandwidth value within a reasonable range. Based on the adaptive bandwidth function, the kernel density is estimated by combining the error samples in the rolling window to obtain the density function. The cumulative distribution function is obtained by integrating the density function. Under the preset confidence level, the upper and lower quantiles are solved according to the cumulative distribution function to obtain the error quantile interval.

[0058] S5. Superimpose the error quantile interval onto the point prediction result, and correct the upper and lower limits according to the physical output boundary of photovoltaic power to output the short-term photovoltaic power prediction interval.

[0059] The error quantile intervals are superimposed onto the point prediction results, and physical boundary constraints are added according to the actual physical meaning of photovoltaic power. The lower bound is corrected to zero and the upper bound is corrected to the rated installed capacity, and the final prediction interval that conforms to the actual physical meaning of photovoltaic power is output.

[0060] Example 2, refer to Figure 3As an embodiment of the present invention, based on the previous embodiment, a method for extracting and identifying load features resistant to voltage changes is provided, including: Furthermore, in S2, the point prediction model is constructed and trained using a historical daily feature sample set, including steps A1-A3: A1. Use a feature extraction network to extract sequence features from the historical daily feature sample set, and input the extracted feature information sequence into a time-series learning network to obtain a prediction vector with time-series features.

[0061] Specifically, the feature extraction network is the Convolutional Neural Network (CNN). CNNs mainly extract sequence features through convolutional layers, pooling layers, and fully connected layers. The core function of a CNN is convolution, expressed as follows: in, These are the original input features; For convolution kernel; For bias terms; To output the feature location; For the convolution kernel row index; Index for the convolution kernel column, On the input feature map, and at the output position And the pixel coordinates corresponding to the convolution kernel index.

[0062] The activation function following convolution can effectively improve the nonlinear adaptability of point prediction models, enabling them to efficiently capture the relationship between input and target variables. (Nonlinear activation function) The expression is: in, This represents the original output feature value.

[0063] The fully connected layers at the back end of a convolutional neural network can output predicted values ​​using activation functions. After training, the point prediction model mines spatial features strongly correlated with the target variable from the dataset, thereby obtaining accurate prediction results. The expression for the linear activation function is as follows: in, Represents the weight matrix. This indicates the activation value of the previous layer. This indicates the bias term.

[0064] This invention introduces a CNN network into the short-term prediction of photovoltaic power on plateaus, leveraging its powerful feature extraction capabilities to fully exploit the temporal features of photovoltaic sequences.

[0065] Specifically, temporal learning networks are also known as long short-term memory networks (LSTM). LSTM networks include input gates, forget gates, output gates, and cell states.

[0066] The input gate is mainly used to control the input data obtained from the historical daily feature sample set at the current time in the network. The amount of information written to the memory cell is as follows: in, The data output by the input gate; For a moment Input data; It is the sigmoid function; The input gate weight matrix; This is the output of the hidden layer at the previous time step; The weight matrix represents the hidden state information in the input gate; For input gate bias, This represents the input gate weight matrix.

[0067] The forget gate is primarily used to determine which information from the cell's previous state should be retained and which should be forgotten. It works by retrieving the hidden state from the previous time step. and the input at the current moment After performing a linear transformation and then processing it with the sigmoid activation function, a value between 0 and 1 is generated. The value is used to characterize the retention rate of cell state information at the current moment, where 0 indicates forgetting and 1 indicates retention. The calculation expressions for the forgetting gate and cell state output are as follows:

[0068] in, Output for the forget gate; For Cellular state at any given moment; for The cell state at time t; tanh is the activation function; Candidate memory cells; Here is the forget gate weight matrix; The weight matrix for the hidden information from the previous time step; The input information weight matrix for the previous time step; Input the information weight matrix for the current time step; Forget gate state bias; Bias for candidate cell states.

[0069] Output gates are primarily used to control the amount of data information output for a new cell state. Specifically, the current cell state... With output gate The hidden layer is jointly determined in Output information at time The output gates and hidden layer output calculation expressions are as follows: in, For a moment The hidden layer output; This is the output of the output gate; This is the output gate weight matrix; The output gate hidden state information weight matrix; This is used to bias the output gate.

[0070] A2. Input the predicted vector into the multi-head attention mechanism layer for weighted processing to obtain the feature vector after attention reconstruction; Furthermore, in A2, the predicted vector is input into the multi-head attention mechanism layer for weighted processing, including A21-A24: A21. Map the prediction vector to the query vector, key vector, and value vector.

[0071] A22. Calculate the similarity between the query vector and the key vector, and normalize the similarity to obtain the attention weight.

[0072] Specifically, the prediction vector is mapped to the query vector. Key vector and value vector First, by and Interaction similarity is calculated to measure the correlation between different parts of the sequence; then, the weights are normalized using the Softmax function.

[0073] A23. Use attention weights to perform a weighted summation of the value vectors to obtain the attention output.

[0074] Based on the weight features, the value vector The attention output is obtained by performing a weighted summation. The output of the attention mechanism is as follows: in, The dimension of the key vector; The query vector and key vector dot product; This is the normalization function.

[0075] A24. Parallel attention calculations are performed in a parallel subspace using multiple attention heads. The outputs of each attention head are then concatenated and fused to obtain the feature vector after attention reconstruction.

[0076] Multi-head attention (MHA) enables parallel attention computation in a subspace, where each attention head has an independent parameter matrix.

[0077] in, For the first Each attention head output; , , All are the first The learnable parameter matrix of each attention head can be used to... , , Mapped to different subspaces; To output the transformation matrix; For the number of heads.

[0078] A3. Input the feature vector into the fully connected layer, use the dynamic time warping distance as a constraint term in the loss function to participate in the parameter update, and output the point prediction result through the point prediction model after training.

[0079] Specifically, in A3, the feature vector is input into the fully connected layer, and the Dynamic Time Warped Distance (DTW) is used as a constraint term in the loss function to participate in parameter updates, including A31-A32: A31. Input the attention-reconstructed feature vector into the fully connected layer to construct the total loss function. With the goal of minimizing the total loss function, update the network parameters of the point prediction model through backpropagation.

[0080] A32. After training, input the feature data of the target day into the point prediction model, and output the point prediction results after forward calculation.

[0081] Specifically, the overall similarity between different sequences is measured through non-linear alignment of sequences along the time axis. The basic principle is as follows: For two given time series and It can build a size of Find a path from the matrix. to The path that minimizes its cumulative distance is: in, Time series middle The value at time; Time series middle The value at time; for and The Euclidean distance; For sequence , DTW distance; For sequence exist Time and Sequence exist Alignment points at specific moments.

[0082] Definition at this time For sequence and The minimum cumulative alignment distance can be obtained by the following recurrence relation:

[0083] The initial conditions are: Here, we consider incorporating DTW into the CNN-LSTM loss function for training. The modified loss function can be expressed as: in, This represents the total training loss of the model. , For MSE loss terms and DTW loss terms; , These are the corresponding weights for the MSE loss term and the DTW loss term.

[0084] Furthermore, in S3, a prediction error sequence is constructed based on the point prediction results and the true values, and a rolling time window is used to locally slice the error sequence to calculate the fuzzy distribution entropy of the error within each window, including steps B1-B3: B1. Extract the point prediction results and corresponding true values ​​at each time point, and construct an absolute error sample by subtracting the values ​​at each point.

[0085] In S3, the fuzzy distribution entropy (FDE) needs to be explained in detail. In the distribution entropy (DE), the original time series is mapped to a number of classes from 1 to 1. A series of integers between these points. Therefore, each element in the sequence needs to be normalized before being assigned an integer using a round function. The round function can be represented as: in, This represents the value of a single time-series element after normalization. Indicates to The value obtained after rounding down.

[0086] Round functions can assign a class to each element in the original sequence, that is, assign a distribution pattern to each embedding vector. In practice, the boundaries between different classes are often fuzzy, as time series are mapped to different categories. To better evaluate the membership of elements, the concept of fuzzy membership is introduced based on distribution entropy, and the fuzzy distribution entropy algorithm (FDE) is proposed. Compared with traditional distribution entropy, fuzzy distribution entropy has lower algorithmic complexity and can better characterize the fluctuation characteristics and nonlinear features of complex signals. For a given length of... time series Its fuzzy distribution entropy can be calculated through the following steps: 1. Use the Normal Cumulative Distribution Function (NCDF) to transform the original sequence Normalize to [0,1] to obtain the normalized sequence. The calculation method is as follows: in, Original sequence Standard deviation; Original sequence The average value.

[0087] 2. Change the sequence Mapped to a sequence according to the following formula :

[0088] in, For sequence The One element, As a class parameter, it indicates what can be assigned to The number of categories of the class.

[0089] 3. In the distribution entropy DE, it is closer to an integer. of Belongs to the Class. Because at the boundary between classes, sequences The element category attribution is ambiguous, therefore a fuzzy membership function is introduced for each category. ,and express For the The membership degree of each category. Where, each They will all be assigned to one or two categories. The membership function has the following characteristics: 1) The beginning of Class 1 and Class 2 The class at the end of a class has no bounds, therefore if Less than 1 or greater than Then it applies to the first and second classes. The membership degree of each class is 1.

[0090] 2) The sum of the membership degrees of different categories is 1, that is: Therefore, the following conclusion can be drawn: in, express For the Membership degree of each category, Indicates to The value obtained after rounding down.

[0091] 3) To maximize the entropy of the sequence, the probability of each category should be equal. Therefore, the fuzzy membership function used should have the same logarithmic cardinality for the random sequence, ensuring that different categories have equal probabilities.

[0092] 4) The support set of the membership function satisfies the following conditions: Among them, parameters The value range is from 0.5 to 1. If If the value is less than 0.5, some elements in the sequence may not be assigned to any category; if If the value is greater than 1, some elements may be assigned to more than two categories. Therefore, to ensure consistency in the analysis, this will be... Let it be 1. Considering the above issues, this paper adopts the following fuzzy function:

[0093]

[0094] in, This represents the time series elements after normalization.

[0095] 4. Therefore, the embedding dimension can be constructed as follows: and time delay is time series : 5. Convert each vector Mapping to distribution patterns based on membership degree In. If and only if Belongs to pattern hour, Category , Category When multiple states are true simultaneously, the t-norm operator can be used for computation. Representing vectors Relative to pattern The membership degree, which is [value]. Relative to each category The product of membership degrees. Its calculation method is as follows: in, express, Indicates the first Starting from a certain time, the interval is... The time step, the first Preprocessed timing values ​​at each location.

[0096] In FDE, each vector The number of categorizable distribution patterns remains at [number]. .

[0097] 6. The probability calculation method for each distribution pattern in the time series is as follows: That is, it belongs to all sequences Distribution pattern The sum of membership degrees and the embedding dimension are The ratio of the total number of embedded signals.

[0098] 7. Finally, the original input sequence can be obtained as follows: The embedding dimension is Number of categories The time delay step is The formula for calculating the fuzzy distribution entropy is: in, Representing distribution patterns The probability, For the corresponding logarithm.

[0099] Since the value range of FDE is 0 to Therefore, after normalization, we can obtain the normalized fuzzy distribution entropy (NFDE): All FDE values ​​mentioned thereafter are normalized values.

[0100] B2. Set a rolling error window on the absolute error sample and slide it along the time axis to divide the full time domain error sequence into several local error subsequences. Specifically, traditional kernel density estimation methods suffer from poor local adaptability and weak nonlinear fitting ability due to the fixed bandwidth. This results in suboptimal interval fitting performance for highly volatile sequences such as photovoltaic data, leading to issues like unreasonable interval width allocation and reduced interval coverage. Since a single FDE index can only roughly characterize the global complexity of the sequence and cannot reflect the local characteristics of the error sequence's evolution over time, a rolling error window is introduced to measure the local FDE level of the error sequence. Based on this, the full-time bandwidth is then evaluated. Adaptive adjustment is implemented to dynamically adjust the model bandwidth for different periods of complexity and volatility, thereby improving its short-term interval forecasting performance in high-altitude photovoltaic multi-meteorological scenarios.

[0101] B3. Calculate the fuzzy distribution entropy of the error subsequence within each rolling window.

[0102] Specifically, an absolute error sample is constructed based on the aforementioned point prediction results: in, The actual value; For point prediction values; This represents the prediction error.

[0103] For time period to Take the length as Rolling error window And calculate the normalized fuzzy distribution entropy within the window. : The larger the value, the higher the complexity of the current local error sequence.

[0104] Furthermore, in step S4, the reference bandwidth is dynamically adjusted based on the fuzzy distribution entropy of each window, and an adaptive bandwidth function is constructed based on the adjusted reference bandwidth, including steps C1-C3: C1. Determine the reference bandwidth based on the statistical characteristics of the error samples.

[0105] Specifically, the baseline bandwidth is selected based on the Silverman rule of thumb: in, The standard deviation of the training error samples; This represents the number of error samples.

[0106] C2. Compare the fuzzy distribution entropy of each window with the fuzzy distribution entropy of the global error sequence, and adjust it using the intensity adjustment coefficient to obtain the adaptive bandwidth.

[0107] C3. Apply broadband boundary constraints to the adjusted adaptive bandwidth to keep the bandwidth value within a reasonable range.

[0108] Specifically, construct an adaptive bandwidth function. And set bandwidth boundary constraints: in, for Time-adaptive bandwidth; The baseline bandwidth; The fuzzy distribution entropy of the current window; The fuzzy distribution entropy of the error sequence over the entire time period; This is the strength adjustment coefficient of the FDE for bandwidth. This indicates the upper limit of adaptive bandwidth. This represents the lower limit of the adaptive bandwidth.

[0109] Furthermore, in S4, kernel density estimation is performed on the error samples within each window based on the adaptive bandwidth function, and the error quantile interval is calculated under a preset confidence level, including steps D1-D4: D1. Based on adaptive bandwidth and combined with error samples within the rolling window, the density function is obtained.

[0110] Specifically, based on the error sample set within the window The adaptive KDE method can be expressed as: Kernel function When using a Gaussian kernel, it can be rewritten as:

[0111] in, For a moment The number of samples participating in kernel density estimation within the rolling window. This represents the standardized input variables of the kernel function. The error variable represents the probability density to be solved. Indicates the first One historical prediction error sample.

[0112] D2. Integrate the density function to obtain the cumulative distribution function.

[0113] Specifically, the cumulative distribution function obtained from the density function is: D3. At the preset confidence level, solve for the upper and lower quantiles corresponding to the confidence level based on the cumulative distribution function.

[0114] Specifically, at a given confidence level The upper and lower quantiles can be obtained as follows: in, express The lower quantile of the time prediction error. express The upper quantile of the time prediction error.

[0115] D4. Based on the upper and lower quantiles, the error interval is obtained.

[0116] Error range .

[0117] Furthermore, in S5, the error quantile intervals are superimposed onto the point prediction results, and upper and lower limits are corrected according to the physical output boundaries of photovoltaic power, outputting the short-term photovoltaic power prediction range, including:

[0118] By superimposing the prediction results of the error quantile points, the final prediction interval is obtained as follows: in, This represents the lower bound of the short-term power forecast range for photovoltaic power. This represents the upper limit of the short-term power forecast range for photovoltaics.

[0119] To ensure that the obtained range conforms to the actual physical meaning of photovoltaic power, physical boundary constraints are added: in, This indicates the maximum installed capacity of the photovoltaic system.

[0120] Example 3 is an embodiment of the present invention. This embodiment provides a plateau photovoltaic short-term point-interval prediction system, including a data acquisition and preprocessing module, a point prediction model construction and training module, an error sequence construction module, an adaptive bandwidth adjustment module, a kernel density estimation and quantile interval calculation module, and an interval synthesis and boundary correction module.

[0121] The data acquisition and preprocessing module acquires historical operational data of plateau photovoltaic power generation and constructs a historical daily feature sample set.

[0122] The point prediction model construction and training module constructs a point prediction model, trains the point prediction model using a historical daily feature sample set, and obtains the point prediction results through the point prediction model.

[0123] The error sequence construction module constructs a prediction error sequence based on the point prediction results and the true values, and uses a rolling time window to locally slice the error sequence and calculate the fuzzy distribution entropy of the error within each window.

[0124] The adaptive bandwidth adjustment module dynamically adjusts the baseline bandwidth based on the fuzzy distribution entropy of each window, and constructs an adaptive bandwidth function based on the adjusted baseline bandwidth.

[0125] The kernel density estimation and quantile interval calculation module performs kernel density estimation on error samples within each window based on an adaptive bandwidth function, and calculates the error quantile interval under a preset confidence level.

[0126] The interval synthesis and boundary correction module superimposes the error quantile intervals onto the point prediction results and corrects them according to the physical output boundary of photovoltaic power, outputting the short-term photovoltaic power prediction interval.

[0127] This embodiment also provides an electronic device applicable to a load feature extraction and identification method resistant to voltage changes, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the load feature extraction and identification method resistant to voltage changes proposed in the above embodiment.

[0128] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a load feature extraction and identification method resistant to voltage changes as proposed in the above embodiments.

[0129] The storage medium proposed in this embodiment and the method for extracting and identifying load features that resist voltage changes proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0130] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0131] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A short-term point-interval prediction method for high-altitude photovoltaic power, characterized in that, include: Acquire historical operational data of photovoltaic power generation on the plateau and construct a historical daily feature sample set; A point prediction model is constructed, and the point prediction model is trained using the historical daily feature sample set. The point prediction result is obtained through the point prediction model. Based on the predicted point results and the actual values, a prediction error sequence is constructed, and a rolling time window is used to locally slice the error sequence to calculate the fuzzy distribution entropy of the error within each window. Based on the fuzzy distribution entropy of each window, the baseline bandwidth is dynamically adjusted. An adaptive bandwidth function is constructed based on the adjusted baseline bandwidth. The kernel density of the error samples in each window is estimated based on the adaptive bandwidth function, and the error quantile interval is calculated under a preset confidence level. The error quantile interval is superimposed on the point prediction result, and the upper and lower limits are corrected according to the physical output boundary of photovoltaic power to output the short-term photovoltaic power prediction interval.

2. The plateau photovoltaic short-term point-interval prediction method according to claim 1, characterized in that: The construction of the point prediction model, which involves training the model using the historical daily feature sample set, includes: The historical daily feature sample set is subjected to sequence feature extraction using a feature extraction network, and the extracted feature information sequence is input into a time-series learning network to obtain a prediction vector with time-series features. The predicted vector is input into the multi-head attention mechanism layer for weighted processing to obtain the feature vector after attention reconstruction. The feature vector is input into a fully connected layer, and the dynamic time warp distance is used as a constraint term in the loss function to participate in parameter updates. After training, the point prediction model outputs the point prediction result.

3. The plateau photovoltaic short-term point-interval prediction method according to claim 2, characterized in that: The step of inputting the predicted vector into the multi-head attention mechanism layer for weighted processing includes: The prediction vector is mapped to a query vector, a key vector, and a value vector; Calculate the similarity between the query vector and the key vector, and normalize the similarity to obtain the attention weight; The attention output result is obtained by weighting and summing the value vector using the attention weights. Parallel attention computation is performed in a parallel subspace using multiple attention heads, and the outputs of each attention head are concatenated and fused to obtain the feature vector after attention reconstruction.

4. The plateau photovoltaic short-term point-interval prediction method according to claim 3, characterized in that: The feature vector is input into a fully connected layer, and the dynamic time-warped distance is used as a constraint term in the loss function to participate in parameter updates, including: The attention-reconstructed feature vector is input into a fully connected layer to construct a total loss function. The network parameters of the point prediction model are updated by backpropagation with the goal of minimizing the total loss function. After training, the feature data of the target day is input into the point prediction model, and the point prediction result is output after forward calculation.

5. The plateau photovoltaic short-term point-interval prediction method according to claim 4, characterized in that: The step of constructing a prediction error sequence based on the point prediction results and the true values, and using a rolling time window to locally slice the error sequence, and calculating the fuzzy distribution entropy of the error within each window, includes: Extract the point prediction results and corresponding true values ​​at each time point, and construct an absolute error sample by subtracting the values ​​at each point. A rolling error window is set on the absolute error sample and slides sequentially along the time axis to divide the full-time error sequence into several local error subsequences; Calculate the fuzzy distribution entropy of the error subsequence within each rolling window.

6. The plateau photovoltaic short-term point-interval prediction method according to claim 5, characterized in that: The process of dynamically adjusting the baseline bandwidth based on the fuzzy distribution entropy of each window, and constructing an adaptive bandwidth function based on the adjusted baseline bandwidth, includes: The baseline bandwidth is determined based on the statistical characteristics of the error samples; The fuzzy distribution entropy of each window is compared with the fuzzy distribution entropy of the global error sequence, and the adaptive bandwidth is obtained by adjusting the intensity adjustment coefficient. Apply broadband boundary constraints to the adjusted adaptive bandwidth to keep the bandwidth value within a reasonable range.

7. The plateau photovoltaic short-term point-interval prediction method according to claim 6, characterized in that: The step of estimating the kernel density of error samples within each window based on an adaptive bandwidth function and calculating the error quantile interval under a preset confidence level includes: Based on adaptive bandwidth and combined with error samples within the rolling window, the density function is obtained; Integrating the density function yields the cumulative distribution function; At a preset confidence level, the upper and lower quantiles corresponding to the confidence level are determined based on the cumulative distribution function. The error range is obtained based on the upper and lower quantiles.

8. A plateau photovoltaic short-term point-interval prediction system, employing the plateau photovoltaic short-term point-interval prediction method as described in any one of claims 1 to 7, characterized in that, include: The module includes a data acquisition and preprocessing module, a point prediction model construction and training module, an error sequence construction module, an adaptive bandwidth adjustment module, a kernel density estimation and quantile interval calculation module, and an interval synthesis and boundary correction module. The data acquisition and preprocessing module acquires historical operating data of plateau photovoltaic power generation and constructs a historical daily feature sample set. The point prediction model construction and training module constructs a point prediction model, trains the point prediction model using the historical daily feature sample set, and obtains point prediction results through the point prediction model. The error sequence construction module constructs a prediction error sequence based on the point prediction results and the true values, and uses a rolling time window to locally slice the error sequence and calculate the fuzzy distribution entropy of the error within each window. The adaptive bandwidth adjustment module dynamically adjusts the reference bandwidth based on the fuzzy distribution entropy of each window, and constructs an adaptive bandwidth function based on the adjusted reference bandwidth. The kernel density estimation and quantile interval calculation module estimates the kernel density of error samples within each window based on an adaptive bandwidth function, and calculates the error quantile interval under a preset confidence level. The interval synthesis and boundary correction module superimposes the error quantile interval onto the point prediction result, and corrects the upper and lower limits according to the physical output boundary of photovoltaic power, outputting the short-term photovoltaic power prediction interval.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the plateau photovoltaic short-term point-interval prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the plateau photovoltaic short-term point-interval prediction method according to any one of claims 1 to 7.

Images