A method, system and device for collaborative optimization of data imputation and power prediction

Abstract

The application discloses a data filling and power prediction collaborative optimization method, system and device, and relates to the fields of artificial intelligence and data processing. The method comprises the following steps: acquiring information data of a target photovoltaic power station; reconstructing the information data based on a time-aware autoencoder to obtain a reconstructed complete power data sequence; wherein, the time stamp is deconstructed and encoded based on a time-aware encoding module to generate a time encoding vector rich in multi-scale periodic information, and the time encoding vector is fused with an embedded vector of the photovoltaic power time sequence; missing values in the fused input sequence are filled and reconstructed based on the autoencoder; the reconstructed complete power data sequence is input into a macro residual transformer for photovoltaic power prediction, and a power prediction value is output. The application can solve the problem of limited prediction performance caused by the separation of data filling and prediction links, and improve the photovoltaic power prediction accuracy.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and data processing, and in particular to a method, system and device for the collaborative optimization of data imputation and power prediction. Background Technology

[0002] With the profound transformation of the global energy structure, the penetration rate of renewable energy, represented by solar energy, in the power system is growing at an unprecedented rate. Photovoltaic power generation, with its abundant resources and clean, pollution-free operation, has become one of the core driving forces for the development of new energy. However, the output power of photovoltaic power generation is highly dependent on meteorological conditions such as solar irradiance and temperature, exhibiting significant intermittency, volatility, and uncertainty. This inherent instability poses a severe challenge to the stable operation of the power grid, power balance, and dispatch planning. Therefore, developing high-precision photovoltaic power prediction technology is of crucial theoretical and practical significance for improving the power grid's capacity to absorb new energy and ensuring the safe and economical operation of the power system.

[0003] Currently, photovoltaic power prediction technology has formed a relatively mature research system, and the mainstream technical solutions can be roughly divided into two categories: physical models and data-driven models. Physical models are based on the physical characteristics of photovoltaic cells and local meteorological parameters, and calculate power generation by establishing complex mathematical and physical equations. Data-driven models, with their high efficiency and flexibility, have been more widely used in the field of short-term power prediction.

[0004] Despite significant advancements in predictive models, a critical bottleneck remains in practical engineering applications: data integrity. Most of the mainstream predictive models are based on an idealized assumption that the input data sequence is complete and continuous. However, in real-world photovoltaic power plant operation and maintenance scenarios, uncontrollable factors such as sensor malfunctions, communication system anomalies, data acquisition equipment maintenance, or extreme weather interference often result in missing data for certain periods.

[0005] This data gap problem impairs the performance of prediction models: during the model training phase, incomplete data sequences disrupt their inherent temporal continuity and correlation, making it difficult for the model to fully learn the periodicity, trend, and randomness of photovoltaic power changes. Summary of the Invention

[0006] The purpose of this application is to provide a method, system, and device for the synergistic optimization of data imputation and power prediction, which can solve the problem of limited prediction performance caused by the separation of data imputation and prediction, and improve the accuracy of photovoltaic power prediction.

[0007] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a collaborative optimization method for data imputation and power prediction, comprising: Acquire information data of the target photovoltaic power plant; the information data includes: a photovoltaic power time series containing missing values ​​and the timestamps corresponding to the photovoltaic power time series; The information data is reconstructed using a time-aware autoencoder to obtain a reconstructed complete power data sequence. The time-aware autoencoder includes a time-aware encoding module and an autoencoder. The time-aware encoding module deconstructs and encodes the timestamps to generate a time-encoded vector rich in multi-scale periodic information, which is then fused with the embedding vector of the photovoltaic power time series to form a fused input sequence. The autoencoder then uses an encoding-decoding process to fill in missing values ​​in the fused input sequence based on a missing marker matrix, outputting the reconstructed complete power data sequence. The missing marker matrix is ​​constructed based on the photovoltaic power time series containing missing values, and is used to characterize whether data corresponding to each time step in the photovoltaic power time series is missing. The reconstructed complete power data sequence is input into the macroscopic residual Transformer for photovoltaic power prediction, and the power prediction value is output. The macroscopic residual Transformer consists of the Transformer of the generative decoder and introduces macroscopic residual connections. The macroscopic residual connections span the Transformer core stack as a global information channel.

[0008] In one embodiment, the timestamp is deconstructed and encoded based on the time-aware encoding module to generate a time-encoded vector rich in multi-scale periodic information, and then fused with the embedding vector of the photovoltaic power time series to form a fused input sequence, specifically including: The timestamp is deconstructed based on the time-aware encoding module to obtain hierarchical time units; the hierarchical time units include: year, month, day, hour, and minute; Encoding conversion is performed based on the hierarchical time units to generate a time-encoded vector rich in multi-scale periodic information; The time encoding vector is fused with the embedding vector of the photovoltaic power time series to form a fused input sequence.

[0009] In one embodiment, encoding conversion is performed based on the hierarchical time units to generate a time-encoded vector rich in multi-scale periodic information, specifically including: Based on the hierarchical time unit, an encoding transformation formula is used to perform encoding conversion, generating a time encoding vector rich in multi-scale periodic information; the expression corresponding to the encoding transformation formula is: ; ; ; ; ; in, To The encoded time-coded vector; This refers to the current day's position within the year; To The encoded time-coded vector; This indicates the current day's position in mid-January; To The encoded time-coded vector; This indicates the current day's position within the week. For hours; In minutes; To The encoded time-coded vector; To The time-coded vector after encoding.

[0010] In one embodiment, based on the autoencoder, missing values ​​in the fused input sequence are filled in and reconstructed according to the missing marker matrix through an encoding-decoding process, and the reconstructed complete power data sequence is output, specifically including: The encoder in the autoencoder compresses the fused input sequence and maps it to a preset low-dimensional hidden layer space to generate a hidden vector. The decoder in the autoencoder receives the latent vector and reconstructs it back into the original data space for filling and reconstruction, outputting the reconstructed complete power data sequence.

[0011] In one embodiment, the encoder in the autoencoder compresses the fused input sequence and maps it to a preset low-dimensional hidden layer space to generate a latent vector. The corresponding processing procedure is as follows: ; in, It is an implicit vector; To compress and map the corresponding processing functions; This refers to the non-linear activation function in the encoder. This is the weight matrix of the encoder in the autoencoder; This is the bias vector of the encoder in the autoencoder; This is for fusing input sequences.

[0012] In one implementation, the expression corresponding to the filling and reconstruction is: ; in, To reconstruct the output data sequence; To fill in the corresponding processing function for reconstruction; This is the non-linear activation function in the decoder; This is the weight matrix of the decoder in the autoencoder; This is the bias vector of the decoder in the autoencoder; It is an implicit vector.

[0013] In one embodiment, before outputting the reconstructed complete power data sequence by filling in missing values ​​in the fused input sequence based on the missing marker matrix through an encoding-decoding process using the autoencoder, the method further includes: The backpropagation algorithm is used to train the parameters of the autoencoder with the goal of minimizing the reconstruction error. The parameters of the autoencoder include: the weight matrix of the encoder, the bias vector of the encoder, the weight matrix of the decoder, and the bias vector of the decoder. The expression for the loss function corresponding to the reconstruction error is: ; in, The loss function corresponding to the reconstruction error; For fusion of input sequences; To reconstruct the output data sequence; To fill in the corresponding processing function for reconstruction; This is for compressing and mapping the corresponding processing functions.

[0014] In one embodiment, the reconstructed complete power data sequence is input into a macroscopic residual Transformer for photovoltaic power prediction, and the output power prediction value specifically includes: The reconstructed complete power data sequence The input is fed into a macroscopic residual Transformer, which performs multi-layer nonlinear transformations to output high-level abstract features. And, the complete power data sequence reconstructed from the linear layer pairs corresponding to the branches connected by macroscopic residuals. Perform feature mapping to generate residual features. ; ; The high-level abstract features and the residual features Element-wise aggregation is performed, and the power prediction value is obtained based on the output layer of the macroscopic residual Transformer. ; ; in, This represents the weight matrix of the linear layer corresponding to the branches connected by the macroscopic residuals. This is the bias vector of the linear layer corresponding to the branch of the macroscopic residual connection.

[0015] Secondly, this application provides a collaborative optimization system for data imputation and power prediction, comprising: The information data acquisition module is used to acquire information data of the target photovoltaic power station; the information data includes: a photovoltaic power time series containing missing values ​​and the timestamps corresponding to the photovoltaic power time series. A sequence reconstruction module is used to reconstruct the information data based on a time-aware autoencoder to obtain a reconstructed complete power data sequence. The time-aware autoencoder includes a time-aware encoding module and an autoencoder. The time-aware encoding module deconstructs and encodes the timestamps to generate a time-encoded vector rich in multi-scale periodic information, which is then fused with the embedding vector of the photovoltaic power time series to form a fused input sequence. The autoencoder then uses an encoding-decoding process to fill in missing values ​​in the fused input sequence based on a missing marker matrix, outputting the reconstructed complete power data sequence. The missing marker matrix is ​​constructed based on the photovoltaic power time series containing missing values, and is used to characterize whether data corresponding to each time step in the photovoltaic power time series is missing. The prediction module is used to input the reconstructed complete power data sequence into the macroscopic residual Transformer to predict photovoltaic power and output the predicted power value. The macroscopic residual Transformer is composed of the Transformer of the generative decoder and introduces macroscopic residual connections. The macroscopic residual connections span the Transformer core stack to serve as a global information channel.

[0016] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned collaborative optimization method for data filling and power prediction.

[0017] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method, system, and device for the collaborative optimization of data imputation and power prediction. The method involves acquiring information data from a target photovoltaic power plant; reconstructing the information data using a time-aware autoencoder to obtain a reconstructed complete power data sequence; deconstructing and encoding the timestamps using a time-aware encoding module to generate a time-coded vector rich in multi-scale periodic information, which is then fused with the embedding vector of the photovoltaic power time series; and imputing missing values ​​in the fused input sequence using the autoencoder to address the limitation in prediction performance caused by the separation of data imputation and prediction. The reconstructed complete power data sequence is then input into a macroscopic residual Transformer for photovoltaic power prediction. By aggregating the reconstructed complete power data sequence, key information is preserved and model robustness is improved, thereby increasing the accuracy of photovoltaic power prediction. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart for a collaborative optimization method for data imputation and power prediction; Figure 2 This is a schematic diagram of the TAEAE-MRT model structure; Figure 3 Flowchart for the implementation of a collaborative optimization method for data imputation and power prediction; Figure 4 A structural diagram of a collaborative optimization system for data completion and power prediction; Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] In one exemplary embodiment, such as Figure 1 As shown, a collaborative optimization method for data imputation and power prediction is provided, including: Step 100: Obtain information data of the target photovoltaic power station. The information data includes: a photovoltaic power time series containing missing values ​​and the timestamps corresponding to the photovoltaic power time series.

[0023] Step 200: Reconstruct the information data based on a time-aware autoencoder to obtain a reconstructed complete power data sequence. The time-aware autoencoder includes a time-aware encoding module and an autoencoder; wherein, based on the time-aware encoding module, the timestamps are deconstructed and encoded to generate a time-encoded vector rich in multi-scale periodic information, and fused with the embedding vector of the photovoltaic power time series to form a fused input sequence; based on the autoencoder, missing values ​​in the fused input sequence are filled and reconstructed according to the missing marker matrix through an encoding-decoding process, and the reconstructed complete power data sequence is output. The missing marker matrix is ​​constructed based on the photovoltaic power time series containing missing values, wherein the missing marker matrix is ​​used to characterize whether the data corresponding to each time step in the photovoltaic power time series is missing.

[0024] In one embodiment, the timestamp is deconstructed and encoded based on the time-aware encoding module to generate a time-encoded vector rich in multi-scale periodic information, and then fused with the embedding vector of the photovoltaic power time series to form a fused input sequence, specifically including: The timestamp is deconstructed based on the time-aware encoding module to obtain hierarchical time units; the hierarchical time units include: year, month, day, hour and minute.

[0025] Encoding conversion is performed based on the hierarchical time units to generate a time-encoded vector rich in multi-scale periodic information.

[0026] Specifically, the encoding transformation based on the hierarchical time units generates a time-encoded vector rich in multi-scale periodic information, including: Based on the hierarchical time unit, an encoding transformation formula is used to perform encoding conversion, generating a time encoding vector rich in multi-scale periodic information; the expression corresponding to the encoding transformation formula is: .

[0027] .

[0028] .

[0029] .

[0030] .

[0031] in, To The encoded time-coded vector; This refers to the current day's position within the year; To The encoded time-coded vector; This indicates the current day's position in mid-January; To The encoded time-coded vector; This indicates the current day's position within the week. For hours; In minutes; To The encoded time-coded vector; To The time-coded vector after encoding.

[0032] The time encoding vector is fused with the embedding vector of the photovoltaic power time series to form a fused input sequence.

[0033] In one embodiment, based on the autoencoder, missing values ​​in the fused input sequence are filled in and reconstructed according to the missing marker matrix through an encoding-decoding process, and the reconstructed complete power data sequence is output, specifically including: The encoder in the autoencoder compresses the fused input sequence and maps it to a preset low-dimensional hidden layer space to generate a hidden vector.

[0034] The decoder in the autoencoder receives the latent vector and reconstructs it back into the original data space for filling and reconstruction, outputting the reconstructed complete power data sequence.

[0035] As an optional implementation, the encoder in the autoencoder compresses the fused input sequence and maps it to a preset low-dimensional hidden layer space to generate hidden vectors. The corresponding processing procedure is as follows: .

[0036] in, It is an implicit vector; To compress and map the corresponding processing functions; This refers to the non-linear activation function in the encoder. This is the weight matrix of the encoder in the autoencoder; This is the bias vector of the encoder in the autoencoder; This is for fusing input sequences.

[0037] The expression for performing fill-in reconstruction is: .

[0038] in, To reconstruct the output data sequence; To fill in the corresponding processing function for reconstruction; This is the non-linear activation function in the decoder; This is the weight matrix of the decoder in the autoencoder; This is the bias vector of the decoder in the autoencoder; It is an implicit vector.

[0039] Before outputting the reconstructed complete power data sequence by filling in missing values ​​in the fused input sequence based on the missing marker matrix through an encoding-decoding process using the autoencoder, the process further includes: The backpropagation algorithm is used to train the parameters of the autoencoder with the goal of minimizing the reconstruction error. The parameters of the autoencoder include: the weight matrix of the encoder, the bias vector of the encoder, the weight matrix of the decoder, and the bias vector of the decoder.

[0040] The expression for the loss function corresponding to the reconstruction error is: .

[0041] in, The loss function corresponding to the reconstruction error; For fusion of input sequences; To reconstruct the output data sequence; To fill in the corresponding processing function for reconstruction; This is for compressing and mapping the corresponding processing functions.

[0042] Step 300: Input the reconstructed complete power data sequence into the macroscopic residual Transformer for photovoltaic power prediction, and output the predicted power value. The macroscopic residual Transformer consists of the Transformer of the generative decoder and introduces macroscopic residual connections; the macroscopic residual connections span the Transformer core stack as a global information channel.

[0043] The reconstructed complete power data sequence is input into the macroscopic residual Transformer for photovoltaic power prediction, and the output power prediction value includes: The reconstructed complete power data sequence The input is fed into a macroscopic residual Transformer, which performs multi-layer nonlinear transformations to output high-level abstract features. And, the complete power data sequence reconstructed from the linear layer pairs corresponding to the branches connected by macroscopic residuals. Perform feature mapping to generate residual features. .

[0044] .

[0045] The high-level abstract features and the residual features Element-wise aggregation is performed, and the power prediction value is obtained based on the output layer of the macroscopic residual Transformer. : .

[0046] This represents the weight matrix of the linear layer corresponding to the branches connected by the macroscopic residuals. This is the bias vector of the linear layer corresponding to the branch of the macroscopic residual connection.

[0047] This application aims to address the limitation in prediction performance caused by the separation of data imputation and prediction processes. First, a Time-Aware Encoding Autoencoder (TAEAE) is constructed to reconstruct photovoltaic power sequences containing missing values. TAEAE comprises a Time-Aware Encoding (TAE) module and an Autoencoder (AE). The TAE module generates a time-encoded vector rich in multi-scale periodic patterns by deconstructing multi-level timestamps (hierarchical time units) such as year, month, day, and hour, and fuses it with the original power sequence (photovoltaic power time series). Subsequently, the autoencoder learns the intrinsic distribution of the fused sequence through an encoding-decoding process, achieving accurate imputation of missing values ​​and outputting a complete and high-fidelity data sequence (the reconstructed complete power data sequence). Next, the reconstructed complete power data sequence is input into a Macro-Residual Transformer (MRT) for power prediction. MRT (Multi-Level Residual Technology) introduces a macroscopic residual connection spanning the entire network core stack, aggregating the model's initial input features with high-level abstract features after multi-layer transformations, thereby preserving key information and improving model robustness. This application significantly improves the accuracy and stability of photovoltaic power prediction by integrating data reconstruction and power prediction into a sequential optimization framework, ensuring the quality of input data before performing deep prediction. Figure 2 This is a schematic diagram of the TAEAE-MRT model structure.

[0048] In other words, data reconstruction and power prediction form a serial processing architecture. First, the data reconstruction step ensures the integrity and high fidelity of the input data. Then, the reconstructed data is used as the input for the power prediction step to improve the accuracy and robustness of the final power prediction value.

[0049] An autoencoder consists of an encoder and a decoder; the encoder is responsible for compressing the fused input sequence and mapping it to a low-dimensional hidden space to generate latent vectors; the decoder is responsible for receiving the latent vectors and reconstructing them back into the original data space to fill in missing values.

[0050] Specifically, TAE is a technique for assigning a unique code to each position in sequence data.

[0051] In sequence modeling tasks, positional encoding is often used to inject sequential information into each position in the sequence in order to enable the model to perceive the sequential relationships between elements. This technique has been widely used in fields such as natural language processing. However, traditional index-based positional encoding methods only reflect the relative or absolute position of data points, ignoring the rich physical and semantic information of the timestamps themselves. This limits the model's ability to capture dependencies and periodic patterns across long periods, as this global information is embedded in the hierarchical timestamps.

[0052] TAE no longer relies on abstract sequence indexes, but instead extracts features directly from physically meaningful timestamps. By deconstructing hierarchical timestamps composed of year, month, day, hour, and minute, TAE can generate a unified encoding vector rich in multi-scale temporal patterns, thus effectively alleviating the performance bottleneck caused by the lack of global information in traditional methods. The specific conversion method is as follows: .

[0053] .

[0054] .

[0055] .

[0056] .

[0057] An AE (Automatic Encoder) is an unsupervised neural network designed to learn efficient representations of data. Its structure mainly consists of an encoder and a decoder.

[0058] The encoder is responsible for processing high-dimensional input data, i.e., fusing the input sequence. Compress and map to a low-dimensional hidden space to generate latent vectors. ,in This process can be represented as: .

[0059] Where d and n are both dimensions.

[0060] The decoder then receives the latent vector. It attempts to reconstruct the data back into the original data space, generating the output, which is the data sequence of the reconstructed output. This process can be represented as: .

[0061] Decoder parameters It is usually different from an encoder.

[0062] The training objective of an autoencoder is to minimize the reconstruction error, i.e. and The differences between them. Their loss function during training is typically as follows: .

[0063] By minimizing the loss function through backpropagation, the model is "forced" to retain as much original information as possible in the low-dimensional latent vector z. After training, the encoder part can be used as a powerful feature extractor or for data dimensionality reduction, while the decoder can be used for generation tasks.

[0064] The core of the macroscopic residual Transformer lies in the introduction of a macroscopic residual connection. The macroscopic residual connection establishes a direct connection path from the initial input of the prediction network to its final output layer. It is used to aggregate the initial input features represented by the reconstructed complete power data sequence with the high-level abstract features after multiple transformations of the prediction network, and then send the aggregated fused features into the final prediction layer.

[0065] Macro-Residual Transformer (MRT) is mainly composed of a Transformer that uses a generative decoder and designs a macro-residual connection.

[0066] As the number of layers in a standard deep Transformer increases, multiple nonlinear transformations can dilute low-level features in the original input, and long paths can hinder gradient backpropagation. MRT improves upon this by using macroscopic residual connections that span the entire Transformer core stack, acting as a global information channel. It establishes a direct path from the initial input to the final output, replacing microscopic residual connections that exist only within sub-layers.

[0067] Specifically, the macroscopic residual connection fuses the encoder's initial input and the decoder's output through a linear layer. Assuming the complete power sequence after missing value imputation (i.e., the initial input of the MRT encoder), which is the reconstructed complete power data sequence, is... After undergoing multiple nonlinear transformations by a Transformer encoder and decoder, the sequence outputs high-level abstract features. The macroscopic residual connection branch passes through an independent linear layer pair. Perform feature mapping to generate residual features. The mathematical expression for this process is: .

[0068] Subsequently, MRT will abstract high-level features and residual characteristics Element-wise aggregation is performed, and the result is fed into the final output layer to obtain the power prediction value. The final output calculation formula is: .

[0069] This design preserves information fidelity, prevents the loss of critical details, and provides a direct shortcut for gradient backpropagation, thereby improving the training stability of deep networks. This mechanism enables the model to effectively fuse high-fidelity raw features with deeply abstract semantic features, thus making more robust predictive decisions.

[0070] As can be seen, this application effectively solves the problem of decreased prediction performance caused by missing data by using an autoencoder to reconstruct photovoltaic power sequences containing missing values ​​with high fidelity, and then feeding the reconstructed complete sequence into the MRT network for accurate prediction, thereby significantly improving the accuracy and robustness of the prediction model.

[0071] like Figure 3 As shown, taking a photovoltaic power station in a certain area as the research object, historical data was collected from January 1, 2022 to December 31, 2023, with a data sampling frequency of one point every 15 minutes. The process includes the following steps: Step S1: Feature selection.

[0072] Before performing the prediction task, to improve the model's prediction accuracy, reduce computational complexity, and avoid overfitting, it is necessary to first construct and screen a feature set highly correlated with photovoltaic power output. In this embodiment, this process first constructs a candidate feature set, which mainly covers two types of information: meteorological features and historical operating features. Meteorological features specifically include total radiation intensity, diffuse radiation intensity, direct radiation intensity, ambient temperature, relative humidity, wind speed, and wind direction. Historical operating features refer to the historical power output sequence of the photovoltaic power station itself, i.e., historical photovoltaic power.

[0073] After constructing a candidate feature set containing eight features, the Pearson correlation coefficient between each feature and the target predicted value was calculated to screen key features. A threshold of 0.4 was set to select features with an absolute correlation coefficient greater than 0.4. Calculations revealed significant correlations between total radiant intensity (0.88), diffuse radiant intensity (0.75), direct radiant intensity (0.81), ambient temperature (0.52), and historical photovoltaic power (0.92). Therefore, the optimal feature subset selected was: {total radiant intensity, diffuse radiant intensity, direct radiant intensity, ambient temperature, historical photovoltaic power}. This set of five features will serve as the formal input for the subsequent prediction model.

[0074] Step S2: Complete the training of TAEAE. This process is broken down into the following steps.

[0075] Convert the timestamp into a location code. Taking a specific time point, June 10, 2023, at 14:30, as an example, the year is 2023, the month is 6, the day of the month is 10, the hour is 14, and the minute is 30. June 10 is the 161st day of the year and the 5th day of the week, calculated with Monday as the 0th day.

[0076] The encoding is converted and then normalized. For example, the mathematical expression for encoding "hour" is as follows: .

[0077] There are 24 hours in a day, and the index is 0-23.

[0078] The following calculations shift it to the interval [-0.5, 0.5]: .

[0079] Perform this operation on all time levels (year, month, day, hour, minute, etc.) to obtain a encoded vector P_time containing multi-scale time information.

[0080] Sliding window data construction. The five features and their corresponding power data selected in step S1 are segmented as input using the sliding window method. In this embodiment, the input window length is set to 24 hours (i.e., 96 data points), and the prediction window length is set to 4 hours (i.e., 16 data points). For example, data from time t_1 to t_96 are used as input X to predict the power values ​​from time t_97 to t_112 as output Y. The sliding window slides across the entire dataset in units of one time step (15 minutes), thereby constructing a large number of {X,Y} sample pairs.

[0081] Data preprocessing and normalization. To simulate data loss in real-world scenarios, 15% of the power data points in the training set were randomly masked. During training, the masked power data points were used to simulate real-world loss scenarios; simultaneously, a loss label matrix was generated based on the masking locations to enable the model to distinguish between actual observation locations and locations to be reconstructed. The training objective of the autoencoder includes not only reconstructing the overall distribution of the complete sequence but also focusing on recovering the power values ​​at the masked locations. Subsequently, max-min normalization was used to scale the values ​​of all features (including unmasked power data) to an interval. The formula is: .

[0082] in, The minimum power data characteristic value, These are the characteristic values ​​of the maximum power data.

[0083] Encoder compression and feature extraction. The preprocessed data sequence is fed into the encoder part of TAEAE. The encoder uses a series of nonlinear transformation layers to compress the high-dimensional input data. Map and compress into a low-dimensional latent vector .

[0084] Decoder reconstruction and missing value imputation. The latent vectors generated by the encoder are then... The latent vectors are fed into the decoder section of TAEAE. The decoder has the opposite structure to the encoder; it receives the latent vectors and attempts to reconstruct them back into the original data dimensions, generating the output sequence. Since the goal of an autoencoder during training is to minimize the input... With output Reconstruction error between: .

[0085] This allows the model to learn the inherent distribution patterns of the data. Therefore, the sequence output by the decoder... It is a complete, high-fidelity data sequence in which the original missing value positions have been reasonably filled in by the network based on learned data patterns. This sequence will serve as the final input for the next prediction task.

[0086] Step S3: Complete the training of MRT.

[0087] Data construction using a sliding window method. In this embodiment, the input window length is set to 24 hours (i.e., 96 data points), and the prediction window length is set to 4 hours (i.e., 16 data points). For example, data from time t_1 to t_96 is used as input X to predict the power values ​​from time t_97 to t_112 as output Y. The sliding window slides across the entire dataset in units of one time step (15 minutes), thereby constructing a large number of {X,Y} sample pairs.

[0088] The encoder performs forward propagation. A historical data window X (96 bytes in length) is fed into a feature extraction network. This network consists of a TAEAE encoder with frozen parameters and an MRT encoder concatenated. First, the historical data window X is fed into the TAEAE encoder trained in step S2 (its weights are frozen and do not participate in gradient updates) to extract its deep temporal features. Then, the latent vector sequence output by the TAEAE encoder is fed into the encoder part of the MRT model. In this embodiment, the MRT encoder consists of two stacked encoder layers, each containing an 8-head self-attention mechanism and a feedforward neural network. The encoder performs deep modeling of the historical sequence, generating a memory representation M rich in contextual information.

[0089] Constructing the generative decoder input and performing direct prediction. This step employs a direct prediction generative decoding paradigm. First, a specific input sequence needs to be constructed for the decoder. The sequence is composed of two parts. The first part is the "teaching items". The first part is typically taken from a later portion of the historical data window X in the time dimension, serving as contextual clues to guide the decoder in generating predictions. The second part is the "target sequence placeholder" ( This is a tensor of the same length as the future real data window Y, with all its elements initialized to 0.

[0090] Join these two parts along the time dimension: .

[0091] Available The memory representation output by the encoder is fed into the decoder. The decoder only needs to perform one forward pass to directly generate the complete prediction sequence for the entire future data window. .

[0092] Loss calculation. The complete prediction sequence generated by the decoder in one go. With real future data window The two are compared to calculate the prediction error between them. Loss function. The mean squared error (MSE) is used to quantify this error.

[0093] Model optimization. Based on the calculated loss L, the gradients of all learnable parameters in the model (mainly the parameters of the MRT part, since the TAEAE encoder is frozen) are calculated using the backpropagation algorithm. Subsequently, the Adam optimizer is used to update the model weights, with a learning rate set to 0.0005 and a batch size of 32. This training process iterates for 100 epochs across the entire dataset and can be combined with an early stopping strategy to prevent overfitting. After training, an MRT model capable of making high-precision, direct, multi-step predictions for the next 4 hours is obtained.

[0094] In one exemplary embodiment, such as Figure 4 As shown, a collaborative optimization system for data imputation and power prediction is provided, comprising: The information data acquisition module is used to acquire information data of the target photovoltaic power station; the information data includes: a photovoltaic power time series containing missing values ​​and the timestamps corresponding to the photovoltaic power time series.

[0095] A sequence reconstruction module is used to reconstruct the information data based on a time-aware autoencoder to obtain a reconstructed complete power data sequence. The time-aware autoencoder includes a time-aware encoding module and an autoencoder. The time-aware encoding module deconstructs and encodes the timestamps to generate a time-encoded vector rich in multi-scale periodic information, which is then fused with the embedding vector of the photovoltaic power time series to form a fused input sequence. The autoencoder then uses an encoding-decoding process to fill in missing values ​​in the fused input sequence based on a missing marker matrix, outputting the reconstructed complete power data sequence. The missing marker matrix is ​​constructed based on the photovoltaic power time series containing missing values, and it characterizes whether data corresponding to each time step in the photovoltaic power time series is missing.

[0096] The prediction module is used to input the reconstructed complete power data sequence into the macroscopic residual Transformer to predict photovoltaic power and output the predicted power value. The macroscopic residual Transformer is composed of the Transformer of the generative decoder and introduces macroscopic residual connections. The macroscopic residual connections span the Transformer core stack to serve as a global information channel.

[0097] In practical applications, the system may include a data reconstruction module and a power prediction module. The data reconstruction module can obtain the photovoltaic power time series containing missing values ​​and the timestamps corresponding to the photovoltaic power time series. In practical applications, the system may include an information data acquisition module, a sequence reconstruction module, and a prediction module. The information data acquisition module acquires photovoltaic power time series containing missing values ​​and their corresponding timestamps. The sequence reconstruction module uses a time-aware encoding module to deconstruct and encode the timestamps, generating a time-encoded vector rich in multi-scale periodic information, and fuses it with the embedding vector of the photovoltaic power time series to form a fused input sequence. The sequence reconstruction module also inputs the fused input sequence to an autoencoder, using an encoding-decoding process to fill in missing values ​​and reconstruct the sequence, outputting a reconstructed complete power data sequence. The prediction module inputs the reconstructed complete power data sequence to a macroscopic residual Transformer to predict future photovoltaic power and outputs a predicted power value.

[0098] The fusion macroscopic residual Transformer configuration in the power prediction module has a macroscopic residual connection, which is used to aggregate the initial input features represented by the reconstructed complete power data sequence with the high-level abstract features inside the prediction network, and use the aggregation result for the final power prediction.

[0099] Specifically, the data reconstruction module encodes timestamps using time-aware encoding to generate time-coded vectors, which are then fused with the power sequence. Missing values ​​are imputed in the fused sequence using an autoencoder. The power prediction module's fusion macroscopic residual Transformer is configured with macroscopic residual connections. These connections aggregate the initial input features represented by the complete power sequence with high-level abstract features within the prediction network, and use the aggregation result for the final power prediction.

[0100] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 5As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data for the co-optimization of data filling and power prediction. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the co-optimization method for data filling and power prediction.

[0101] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0102] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0103] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0104] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0105] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0106] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0107] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logic devices, etc., and are not limited to these.

[0108] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0109] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for the coordinated optimization of data imputation and power prediction, characterized in that, include: Acquire information data of the target photovoltaic power plant; the information data includes: a photovoltaic power time series containing missing values ​​and the timestamps corresponding to the photovoltaic power time series; The information data is reconstructed using a time-aware autoencoder to obtain a reconstructed complete power data sequence. The time-aware autoencoder includes a time-aware encoding module and an autoencoder. The time-aware encoding module deconstructs and encodes the timestamps to generate a time-encoded vector rich in multi-scale periodic information, which is then fused with the embedding vector of the photovoltaic power time series to form a fused input sequence. The autoencoder then uses an encoding-decoding process to fill in missing values ​​in the fused input sequence based on a missing marker matrix, outputting the reconstructed complete power data sequence. The missing marker matrix is ​​constructed based on the photovoltaic power time series containing missing values, and is used to characterize whether data corresponding to each time step in the photovoltaic power time series is missing. The reconstructed complete power data sequence is input into the macroscopic residual Transformer for photovoltaic power prediction, and the power prediction value is output. The macroscopic residual Transformer consists of the Transformer of the generative decoder and introduces macroscopic residual connections. The macroscopic residual connections span the Transformer core stack as a global information channel.

2. The method for synergistic optimization of data imputation and power prediction according to claim 1, characterized in that, The time-aware encoding module deconstructs and encodes the timestamp to generate a time-encoded vector rich in multi-scale periodic information, which is then fused with the embedding vector of the photovoltaic power time series to form a fused input sequence, specifically including: The timestamp is deconstructed based on the time-aware encoding module to obtain hierarchical time units; the hierarchical time units include: year, month, day, hour, and minute; Encoding conversion is performed based on the hierarchical time units to generate a time-encoded vector rich in multi-scale periodic information; The time encoding vector is fused with the embedding vector of the photovoltaic power time series to form a fused input sequence.

3. The method for synergistic optimization of data imputation and power prediction according to claim 2, characterized in that, Encoding transformation is performed based on the hierarchical time units to generate a time-encoded vector rich in multi-scale periodic information, specifically including: Based on the hierarchical time unit, an encoding transformation formula is used to perform encoding conversion, generating a time encoding vector rich in multi-scale periodic information; the expression corresponding to the encoding transformation formula is: ； ； ； ； ； in, To The encoded time-coded vector; This refers to the current day's position within the year; To The encoded time-coded vector; This indicates the current day's position in mid-January; To The encoded time-coded vector; This indicates the current day's position within the week. For hours; In minutes; To The encoded time-coded vector; To The time-coded vector after encoding.

4. The method for synergistic optimization of data imputation and power prediction according to claim 1, characterized in that, Based on the autoencoder, the missing values ​​in the fused input sequence are filled in and reconstructed according to the missing marker matrix through an encoding-decoding process, and the reconstructed complete power data sequence is output, specifically including: The encoder in the autoencoder compresses the fused input sequence and maps it to a preset low-dimensional hidden layer space to generate a hidden vector. The decoder in the autoencoder receives the latent vector and reconstructs it back into the original data space for filling and reconstruction, outputting the reconstructed complete power data sequence.

5. The method for coordinated optimization of data imputation and power prediction according to claim 4, characterized in that, The encoder in the autoencoder compresses the fused input sequence and maps it to a preset low-dimensional hidden layer space to generate a hidden vector. The corresponding processing steps are as follows: ； in, It is an implicit vector; To compress and map the corresponding processing functions; This refers to the non-linear activation function in the encoder. This is the weight matrix of the encoder in the autoencoder; This is the bias vector of the encoder in the autoencoder; This is for fusing input sequences.

6. The method for coordinated optimization of data imputation and power prediction according to claim 4, characterized in that, The expression for performing fill-in reconstruction is: ； in, To reconstruct the output data sequence; To fill in the corresponding processing function for reconstruction; This is the non-linear activation function in the decoder; This is the weight matrix of the decoder in the autoencoder; This is the bias vector of the decoder in the autoencoder; It is an implicit vector.

7. The method for synergistic optimization of data imputation and power prediction according to claim 1, characterized in that, Before outputting the reconstructed complete power data sequence by filling in missing values ​​in the fused input sequence based on the missing marker matrix through an encoding-decoding process using the autoencoder, the process further includes: The backpropagation algorithm is used to train the parameters of the autoencoder with the goal of minimizing the reconstruction error. The parameters of the autoencoder include: the weight matrix of the encoder, the bias vector of the encoder, the weight matrix of the decoder, and the bias vector of the decoder. The expression for the loss function corresponding to the reconstruction error is: ； in, The loss function corresponding to the reconstruction error; For fusion of input sequences; To reconstruct the output data sequence; To fill in the corresponding processing function for reconstruction; This is for compressing and mapping the corresponding processing functions.

8. The method for collaborative optimization of data imputation and power prediction according to claim 1, characterized in that, The reconstructed complete power data sequence is input into the macroscopic residual Transformer for photovoltaic power prediction, and the output power prediction value includes: The reconstructed complete power data sequence The input is fed into a macroscopic residual Transformer, which performs multi-layer nonlinear transformations to output high-level abstract features. And, the complete power data sequence reconstructed from the linear layer pairs corresponding to the branches connected by macroscopic residuals. Perform feature mapping to generate residual features. ; ； The high-level abstract features and the residual features Element-wise aggregation is performed, and the power prediction value is obtained based on the output layer of the macroscopic residual Transformer. ; ； in, This represents the weight matrix of the linear layer corresponding to the branches connected by the macroscopic residuals. This is the bias vector of the linear layer corresponding to the branch of the macroscopic residual connection.

9. A collaborative optimization system for data imputation and power prediction, characterized in that, include: The information data acquisition module is used to acquire information data of the target photovoltaic power station; the information data includes: a photovoltaic power time series containing missing values ​​and the timestamps corresponding to the photovoltaic power time series. A sequence reconstruction module is used to reconstruct the information data based on a time-aware autoencoder to obtain a reconstructed complete power data sequence. The time-aware autoencoder includes a time-aware encoding module and an autoencoder. The time-aware encoding module deconstructs and encodes the timestamps to generate a time-encoded vector rich in multi-scale periodic information, which is then fused with the embedding vector of the photovoltaic power time series to form a fused input sequence. The autoencoder then uses an encoding-decoding process to fill in missing values ​​in the fused input sequence based on a missing marker matrix, outputting the reconstructed complete power data sequence. The missing marker matrix is ​​constructed based on the photovoltaic power time series containing missing values, and is used to characterize whether data corresponding to each time step in the photovoltaic power time series is missing. The prediction module is used to input the reconstructed complete power data sequence into the macroscopic residual Transformer to predict photovoltaic power and output the predicted power value. The macroscopic residual Transformer is composed of the Transformer of the generative decoder and introduces macroscopic residual connections. The macroscopic residual connections span the Transformer core stack to serve as a global information channel.

10. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the collaborative optimization method for data imputation and power prediction according to any one of claims 1-8.

Images