A deep learning-based data time period missing value imputation method and system

Abstract

The application relates to the technical field of photovoltaic power generation power prediction, and discloses a data time period missing value interpolation method and system based on deep learning, which comprises the following steps: acquiring power generation power data and meteorological data, identifying a data missing time period, and acquiring context time period data; searching for photovoltaic power stations in a space region where a target photovoltaic power station is located, and screening to form a photovoltaic candidate power station set; adopting a self-encoder based on a time sequence convolution network to extract features of a power generation power curve, so as to obtain a time sequence feature vector; calculating the similarity score of the target photovoltaic power station and the photovoltaic candidate power station, and constituting a dynamic space reference power station set; according to the weight, weighting and averaging the actual power values of each reference power station in the missing time period; combining the meteorological data to calculate the upper limit of the theoretical power generation power of the target photovoltaic power station, applying physical constraints to the preliminary completed value, and taking the minimum value as the final completed value. The application realizes high-precision and self-adaptive filling of the missing data in the time period.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, specifically to a method and system for interpolating missing values ​​in photovoltaic power generation data over time periods based on deep learning. Background Technology

[0002] In the specific task of photovoltaic (PV) power generation forecasting, PV data is not a typical time series; it exhibits strong and complex spatiotemporal dependencies. In the data preprocessing stage of the forecasting task, data quality is fundamental to predictive performance. Actual operational data from PV power plants often suffers from varying degrees of data loss due to sensor malfunctions, communication interruptions, and equipment shutdowns caused by severe weather (such as sandstorms, heavy rain, and heavy snow).

[0003] Current technologies generally employ linear interpolation, mean imputation, or simple machine learning models for missing value completion. However, these methods exhibit significant limitations when faced with long-term, non-random, and patterned missing values ​​caused by complex and variable weather (such as continuous overcast or rainy weather, rapidly moving clouds, and sudden weather events). For example, when several consecutive hours of overcast or rainy weather result in zero or extremely low power generation for an extended period, simple interpolation methods fail to capture this pattern of missing values ​​driven by specific weather conditions. This leads to distorted imputed data, introducing false information and severely misleading subsequent model predictions, causing systematic prediction bias. Therefore, there is an urgent need to develop a method for effectively handling numerical missing values ​​over long periods, providing a more robust technical guarantee for the training data of subsequent photovoltaic power generation prediction. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides a data time-segmentation missing value interpolation method and system based on deep learning. This method enables high-precision, adaptive filling of missing data in time periods with specific distribution characteristics in the field of photovoltaic power generation prediction. It can be extended to other renewable energy power generation prediction or meteorological time series data missing value processing with periodicity and volatility.

[0005] To achieve the aforementioned objectives, this invention employs a data time-segmentation missing value interpolation method and system based on deep learning. The specific technical solution is as follows:

[0006] In a first aspect, this invention proposes a method for imputing missing values ​​in data time periods based on deep learning, comprising the following steps:

[0007] Acquire power generation data and meteorological data of the target photovoltaic power station, identify the time period with missing data for the target photovoltaic power station, and obtain the contextual time period data for the time period with missing data;

[0008] Search for photovoltaic power stations within the spatial area where the target photovoltaic power station is located, and filter them to form a set of candidate photovoltaic power stations based on the installed capacity and geographical location information of each photovoltaic power station;

[0009] An autoencoder based on a temporal convolutional network is used to extract features from the power generation curves of the photovoltaic candidate power station set and the target photovoltaic power station to obtain a temporal feature vector;

[0010] The similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period is calculated based on the time series feature vector. The Top-K candidate power stations are selected based on the similarity score to form a dynamic spatial reference power station set.

[0011] The similarity scores of each reference power station are normalized into weights, and the actual power values ​​of each reference power station are weighted and averaged at each time point within the missing period to obtain the preliminary completion value.

[0012] Based on meteorological data, the theoretical power generation limit of the target photovoltaic power station is calculated. Physical constraints are applied to the preliminary supplementary value, and the minimum value is taken as the final supplementary value; including:

[0013] Using meteorological data from the target power plant and measured irradiance data from meteorological stations for the missing time period, the theoretical upper limit of power generation is calculated through physical modeling. ;

[0014] Apply physical constraints to the initial completion values ​​at each time point to ensure that the final completion values ​​meet the following conditions. ;

[0015] For each time point within the missing period, the actual power values ​​of each reference power station are weighted and averaged to obtain the preliminary complete value, specifically:

[0016] in, This represents the initial completed value for the target photovoltaic power plant at time point t. Let i be the weight of the i-th photovoltaic power station in the dynamic spatial reference power station set. Let be the actual power value of the i-th photovoltaic power station in the dynamic spatial reference power station set at time t.

[0017] Optionally, a set of candidate photovoltaic power plants can be formed based on the installed capacity and geographical location information of each photovoltaic power plant, including:

[0018] The primary criterion is to ensure the consistency of installed capacity, and photovoltaic power plants with capacities different from the target photovoltaic power plant will be eliminated.

[0019] Based on the time period of missing data for the target photovoltaic power station, photovoltaic power stations whose data is unavailable in the current time period are further excluded.

[0020] Optionally, the autoencoder based on the temporal convolutional network includes an encoder, a bottleneck layer, and a decoder; wherein the encoder extracts temporal features through causal convolution and dilated convolution, the bottleneck layer forces the learning of essential feature representations, and the decoder symmetrically reconstructs the sequence.

[0021] Optionally, the encoder includes multiple stacked temporal convolutional network modules. Within each temporal convolutional network module, a causal convolutional layer with an asymptotic dilation rate is used to achieve asymptotic downsampling of the sequence length. Between adjacent modules, a convolutional or pooling layer with a stride > 1 is used to achieve asymptotic downsampling of the sequence length.

[0022] Optionally, the autoencoder based on the temporal convolutional network is trained using a self-supervised learning approach, with the input sequence itself as the reconstruction target, and the Adam optimizer is used to minimize the reconstruction error.

[0023] Optionally, the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period is calculated based on the time-series feature vector using at least one of cosine similarity or Manhattan distance, and the measurement method is adaptively selected based on the volatility of the power generation curve.

[0024] Optionally, for the stable region of the power generation curve, cosine similarity is used to calculate the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period;

[0025] For the fluctuating regions of the power generation curve, the Manhattan distance is used to calculate the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context of the time period.

[0026] Secondly, this invention proposes a data time-segmentation missing value interpolation system based on deep learning, applying the value interpolation method described above, including:

[0027] The data acquisition module is used to acquire the power generation data and meteorological data of the target photovoltaic power station, identify the time period when the target photovoltaic power station has missing data, and acquire the context time period data of the time period when the data is missing.

[0028] The candidate power station processing module is used to search for photovoltaic power stations within the spatial area where the target photovoltaic power station is located, and to filter and form a set of photovoltaic candidate power stations based on the installed capacity and geographical location information of each photovoltaic power station.

[0029] The temporal feature extraction module is used to extract features from the power generation curves of the photovoltaic candidate power station set and the target photovoltaic power station using an autoencoder based on a temporal convolutional network, and obtain a temporal feature vector.

[0030] The dynamic reference set construction module is used to calculate the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period based on the time-series feature vector, and select the Top-K candidate power stations based on the similarity score to form a dynamic spatial reference power station set;

[0031] The dynamic weighted interpolation module is used to normalize the similarity scores of each reference power station into weights, and to calculate the weighted average of the actual power values ​​of each reference power station at each time point within the missing time period to obtain the preliminary complete value.

[0032] The completion output module is used to calculate the theoretical power generation limit of the target photovoltaic power station by combining meteorological data, apply physical constraints to the preliminary completion value, and take the minimum value as the final completion value.

[0033] The present invention has the following beneficial effects:

[0034] This invention overcomes the inherent defects of traditional methods, avoiding data distortion caused by using fixed reference points or simple curve matching, and effectively avoiding noise interference in similarity calculation. It provides a reliable, efficient, and interpretable technical foundation for intelligent operation and maintenance of photovoltaic systems. Its dynamic optimization mechanism can lock the most relevant reference data for each missing time period, thereby fundamentally improving the quality of data completion. This method is not only applicable to parameter repair of a single photovoltaic site, but can also be extended to data quality improvement of photovoltaic clusters and distributed photovoltaic networks, providing reliable technical support for intelligent operation and maintenance of large-scale photovoltaic systems. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the missing value interpolation method of the present invention;

[0036] Figure 2 This is a time series diagram showing the missing time periods of the power generation of the target photovoltaic power plant in this invention;

[0037] Figure 3 This is a diagram of the autoencoder network structure based on a temporal convolutional network of the present invention;

[0038] Figure 4 This is a comparison chart of the normalized similarity calculation results of the present invention;

[0039] Figure 5 This is a graph showing the results of filling in missing values ​​for the time period in this invention.

[0040] Figure 6 This is a comparison chart showing the results of filling missing values ​​in different time periods using the algorithms of this invention. Detailed Implementation

[0041] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0042] This invention proposes a method for filling missing time period data in photovoltaic power generation forecasting. The idea behind this method is that, within the non-missing time period [no_t_start, no_t_end], if the curve of a neighboring photovoltaic station B is similar to the curve of the target photovoltaic station T, then within the missing time period [t_start, t_end], the value of the neighboring station B in this missing time period can be used to calculate the "missing time period value of the target station T". Specifically, a method for filling missing time period values ​​in photovoltaic power generation data based on dynamic similarity weighted interpolation is disclosed.

[0043] The core of this method lies in using deep learning technology to dynamically find the most relevant reference information for each period of missing data and intelligently fill in the gaps. Specifically, firstly, a pre-trained autoencoder based on a temporal convolutional network is used to learn the deep essential features of photovoltaic power generation or other curves from historical data, transforming complex time-series data into low-dimensional feature vectors; then, when the target power station has missing data, the similarity (such as cosine similarity) between its feature vectors and those of other reference power stations in the region is calculated, and the K most similar reference power stations are dynamically selected accordingly.

[0044] Calculating similarity using feature vectors relies on learning the physical laws of photovoltaics. Compared to traditional similarity calculation methods, this method, based on physical laws rather than surface numerical values, has strong business significance and can intelligently filter noise and redundancy. Finally, weights are assigned based on the calculated similarity, and a weighted average is performed on the power generation data of the missing time periods of the reference power station to generate the missing time period values ​​for the target power station.

[0045] This invention overcomes the inherent defects of traditional methods, avoiding data distortion caused by using fixed reference points or simple curve matching, and effectively avoiding noise interference in similarity calculation. It provides a reliable, efficient, and interpretable technical foundation for intelligent operation and maintenance of photovoltaic systems. Its dynamic optimization mechanism can lock the most relevant reference data for each missing time period, thereby fundamentally improving the quality of data completion. This method is not only applicable to parameter repair of a single photovoltaic site, but can also be extended to data quality improvement of photovoltaic clusters and distributed photovoltaic networks, providing reliable technical support for intelligent operation and maintenance of large-scale photovoltaic systems.

[0046] like Figure 1 As shown, an embodiment of the present invention provides a data time-time missing value interpolation method based on deep learning, comprising the following steps S1 to S6:

[0047] S1. Obtain the power generation data and meteorological data of the target photovoltaic power station, identify the time period when the data of the target photovoltaic power station is missing, and obtain the context time period data of the time period when the data is missing;

[0048] In an optional embodiment of the present invention, step S1 first locates the missing value type, determines which missing mechanism the missing values ​​in the dataset that needs to be filled in for photovoltaic site T belong to, and identifies their positions; then determines the missing time period and obtains context data. The power generation data of the target power station T is missing in the time period [t_start, t_end], and the context data of the missing time period of its power generation curve is [no_t_start, no_t_end].

[0049] This embodiment obtains the power generation data or meteorological data of the target power station that needs to have missing values ​​filled, establishes a unified time series database, and sets the time resolution to 15 minutes per sample point.

[0050] Identify the missing data period for the target power plant: T_missing = [t_start, t_end]. Based on photovoltaic characteristics (nighttime zeroing, curtailment, etc.), comprehensively distinguish the missing data type: random point missing and time-period missing. Random point missing data refers to scattered, isolated missing points; time-period missing data refers to continuous, blocky missing intervals. This data is identified as a continuous time-period missing data type, and the context information of the missing period is recorded. The power generation curve and the missing data period are shown below. Figure 2 As shown, there are 96 sampling points per day, and data was collected for 4 days, totaling 384 time sampling points. The missing data is in the time period [144, 159].

[0051] S2. Search for photovoltaic power stations within the spatial area where the target photovoltaic power station is located, and filter them to form a set of photovoltaic candidate power stations based on the installed capacity and geographical location information of each photovoltaic power station;

[0052] In an optional embodiment of the present invention, step S2, which involves filtering and forming a set of photovoltaic candidate power stations based on the installed capacity and geographical location information of each photovoltaic power station, includes:

[0053] The primary criterion is to ensure the consistency of installed capacity, and photovoltaic power plants with capacities different from the target photovoltaic power plant will be eliminated.

[0054] Based on the time period of missing data for the target photovoltaic power station, photovoltaic power stations whose data is unavailable in the current time period are further excluded.

[0055] In this embodiment, candidate power station pool preprocessing is performed in step S2. Photovoltaic power stations within the spatial area are searched, and power stations with complete and usable data during the missing time period [t_start, t_end] are selected to form the available candidate set for this completion task.

[0056] Taking a photovoltaic power station with missing data periods as an example, multiple candidate reference power stations are searched within the area.

[0057] First, a static candidate power station reference pool was established for nine sites near the photovoltaic power station. Second, based on the installed capacity and geographical location information of each site, as shown in Table 1, a set of photovoltaic candidate power stations was formed. The selection principles were as follows: (1) The consistency of installed capacity was the primary criterion, and photovoltaic stations A, E, and F with different capacities from the target power station were eliminated. (2) Dynamic availability filtering: When processing each specific missing event, power stations whose data was unavailable at that time were automatically excluded from the static pool. (3) Final candidate list: After the above selection, the candidate reference power stations available for this calculation were determined to be: stations B, C, D, G, H, and I. Finally, the complete power generation data of the above six candidate reference power stations during the target period (a total of four days, including the missing period [t_start, t_end]) was obtained. The specific data of the candidate power stations are shown in Table 2.

[0058] Table 1. Photovoltaic Site Installed Capacity and Latitude / Longitude

[0059]

[0060] Table 2. Data Table of Candidate Photovoltaic Power Plant Pools

[0061]

[0062] S3. An autoencoder based on a temporal convolutional network is used to extract features from the power generation curves of the photovoltaic candidate power station set and the target photovoltaic power station to obtain a temporal feature vector.

[0063] In an optional embodiment of the present invention, the autoencoder based on a temporal convolutional network used in step S3 includes an encoder, a bottleneck layer, and a decoder; wherein the encoder extracts temporal features through causal convolution and dilated convolution, the bottleneck layer forces the learning of essential feature representations, and the decoder symmetrically reconstructs the sequence.

[0064] The encoder includes multiple stacked temporal convolutional network modules. Within each temporal convolutional network module, a causal convolutional layer with an asymptotic dilation rate is used to achieve asymptotic downsampling of the sequence length. Between adjacent modules, a convolutional or pooling layer with a stride > 1 is used to achieve asymptotic downsampling of the sequence length.

[0065] The autoencoder based on temporal convolutional networks is trained using a self-supervised learning approach, with the input sequence itself as the reconstruction target, and uses the Adam optimizer to minimize the reconstruction error.

[0066] In this embodiment, feature vectors are output in step S3. The curves of photovoltaic power stations in the candidate power station pool are used as inputs to an autoencoder based on a temporal convolutional network. The autoencoder based on the temporal convolutional network outputs the curve feature vectors of each photovoltaic power station.

[0067] Figure 3This embodiment designs an autoencoder structure based on a temporal convolutional network. The encoder extracts multi-scale temporal features through causal convolution and dilated convolution, the decoder symmetrically reconstructs the sequence, and the intermediate bottleneck layer forces the learning of low-dimensional feature representations. The construction of the autoencoder based on the temporal convolutional network consists of three parts: the encoder, the intermediate bottleneck layer, and the decoder. Here, "+" indicates that residual mapping + original information = complete input-output mapping. The TCN autoencoder, composed of the encoder, the information bottleneck layer, and the decoder, aims to extract important essential feature vectors from the temporal sequences of photovoltaic sites, providing a foundation for subsequent calculations of temporal sequence similarity between photovoltaic sites.

[0068] 1. The encoder section, primarily used to extract essential features from time-series sequences, features a relatively deep network structure, including four causal convolutional layers. To address the vanishing gradient problem that may occur in deep networks, the model introduces a residual structure. The idea is that directly learning the complete mapping H(x) of the input vector x after passing through the TCN network is difficult, while learning the residual mapping F(x) = H(x) − x is relatively easy. This approach simplifies the network's data complexity, enhances its training stability and feature extraction capabilities. The plus sign "+" stands for "residual addition." This design allows the model to more effectively capture key temporal features in the sequence, forming representative feature vectors for subsequent decoding and similarity analysis.

[0069] 2. The complete feature map H(x) extracted by the encoder is sent to the information bottleneck layer for compression and purification of key information. This process aims to retain the core temporal pattern of the sequence, remove redundant information, and finally output a compressed feature vector Z with reduced dimensions and high abstraction.

[0070] 3. This vector Z is then used as the input to the decoder. The decoder adopts a mirror-symmetric structure with the encoder, and its main function is to reconstruct the input vector. It performs upsampling through transposed causal dilated convolutional layers, and gradually restores the dimension and details of the data through other causal convolutional layers, finally reconstructing the output vector y.

[0071] The goal of model training is to minimize the error (e.g., mean squared error) between the reconstructed output y and the original input x. Through iterative optimization, when the reconstruction error converges to a minimum, it means that the encoder-bottleneck layer has learned to extract the most representative and essential feature representation from the original sequence. The optimal feature vector Z1 obtained at this point is the effective feature vector for the essential information of the photovoltaic power plant time series.

[0072] In subsequent calculations of similarity between sites using this essential feature vector Z1, the similarity between different photovoltaic sequences can be accurately and stably reflected.

[0073] Each temporal convolutional network module contains a one-dimensional convolutional layer, a ReLU activation function, layer normalization, and a Dropout layer. It also mitigates the vanishing gradient problem through residual connections, enabling the network to be trained to be deeper.

[0074] The encoder consists of multiple stacked temporal convolutional network modules, employing a progressive dilation rate design (1, 2, 4, 8 structure). Progressive downsampling of the sequence length is achieved through convolutional layers with a stride greater than 1 (L → L / 2 → L / 4 → L / 8), ultimately forming a compressed sequence with a shorter length but richer features as the information bottleneck.

[0075] The encoder consists of multiple stacked temporal convolutional network modules. Within each temporal convolutional network module, causal convolutional layers with progressively increasing dilation rates (e.g., 1, 2, 4, 8) are used to expand the receptive field while maintaining the sequence length. Between adjacent temporal convolutional network modules, progressive downsampling of the sequence length is achieved through convolutional or pooling layers with a stride > 1 (e.g., L → L / 2 → L / 4 → L / 8). Finally, the encoder outputs a significantly shortened representation that contains global information about the input sequence; this representation can serve as an information bottleneck for subsequent processing.

[0076] The intermediate bottleneck layer acts as an information bottleneck, forcing the network to learn the essential feature representation of the input sequence and store the essential features of the data. These features are used for sequence reconstruction and can also be extracted to support subsequent similarity calculation and downstream analysis tasks.

[0077] The decoder employs a symmetrical architecture to the encoder, progressively upsampling and reconstructing the compressed feature sequence into an output sequence of the original length through transposed convolutional layers. The decoding process utilizes the same number of filters and layers as the encoder to ensure information flow symmetry and reconstruction quality. The final output layer uses a linear activation function to meet the requirements of regression tasks in sequence reconstruction.

[0078] The entire network forces the learning of the essential features of photovoltaic curve data through this symmetrical encoder-decoder structure. The autoencoder structure based on a temporal convolutional network is as follows: Figure 3 As shown.

[0079] The autoencoder based on temporal convolutional networks is trained using a self-supervised learning approach. Its core idea is to use the input sequence itself as the training objective. By minimizing the reconstruction error, the autoencoder model can learn the intrinsic temporal dependence and variation patterns of the photovoltaic power output curve from the data, without relying on manually labeled data. The Adam optimizer, known for its stability and fast convergence in deep learning tasks and suitability for handling sequential data and non-stationary objectives, is used uniformly. The initial learning rate is set to 0.001. Through this self-supervised training framework, the model can learn essential features such as "typical daily cycles," "seasonal cycles," "weather-sensitive characteristics," and "transient patterns" from a large amount of unlabeled historical photovoltaic data. The resulting feature representation (encoder output) can then be used as input for subsequent bottleneck layers. The encoder input is a standardized fixed-length power generation sequence, such as [batch, 1, 96]. The encoder output is a high-dimensional intermediate feature map, which contains rich multi-scale information but has a lot of redundancy. This input is sent to the bottleneck layer, and the bottleneck layer output is such as [batch, 32]. Each dimension corresponds to the specific physical meaning of the photovoltaic curve. The bottleneck layer output is a low-dimensional feature vector after information compression and semantic encoding.

[0080] The output of the bottleneck layer is fed into the decoder for reconstruction. The reconstructed curve is compared with the original curve for error calculation (e.g., RMSE). The error gradient is then propagated back through backpropagation to the decoder, bottleneck layer, and encoder, driving simultaneous optimization of these three parts. The encoder improves by extracting more essential features, the bottleneck layer adjusts feature combinations to retain key information, and the decoder learns to reconstruct the curve from the features. Through continuous iteration, the autoencoder network based on a temporal convolutional network gradually learns to accurately encode the core patterns of the photovoltaic curve, such as local and daily cycles and equipment status, using low-dimensional features, minimizing the reconstruction error. Simultaneously, the refined features become ideal input for downstream similarity calculations.

[0081] Once the autoencoder based on a temporal convolutional network converges during training and the reconstruction error reaches a sufficiently small level, model training is essentially complete. At this point, the power generation sequences of each photovoltaic power station, after standardized preprocessing, are input into the model. The encoder extracts temporal features, which are then compressed into low-dimensional feature representations by a bottleneck layer. To ensure consistent feature dimensions across all input sequences, the original data needs to be uniformly adjusted to a fixed length using a sliding window or padding method. After this standardized preprocessing, the simplified feature vector output by the bottleneck layer fully captures the essential temporal dependencies and change patterns in the input sequences and can be directly used for subsequent sequence similarity calculations and analysis.

[0082] S4. Calculate the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period based on the time series feature vector, and select the Top-K candidate power stations based on the similarity score to form a dynamic spatial reference power station set.

[0083] In an optional embodiment of the present invention, step S4 calculates the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period based on the time-series feature vector, using at least one of cosine similarity or Manhattan distance, and adaptively selects the measurement method based on the volatility of the power generation curve.

[0084] For the stable region of the power generation curve, cosine similarity is used to calculate the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period;

[0085] For the fluctuating regions of the power generation curve, the Manhattan distance is used to calculate the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context of the time period.

[0086] In this embodiment, context similarity calculation and dynamic reference set construction are performed in step S4. First, the feature vectors of the curves of each candidate power station are obtained using an autoencoder based on a temporal convolutional network. The similarity between the curve feature vector of each candidate power station and the curve feature of the target power station T in the context of the missing value is calculated, and the similarity score between the curves of the candidate power station and the curve of the target power station is quantified. Then, based on the similarity score, the K most similar power stations are selected to form a dynamic spatial reference power station set specifically for this missing value event.

[0087] This embodiment uses cosine similarity as the primary metric in the feature vector space. This method focuses on the directional similarity of feature vectors rather than absolute distance and is insensitive to changes in feature scale. Manhattan distance similarity is also utilized. An adaptive similarity strategy is adopted for different types of photovoltaic curve characteristics. The feature vector of the target power generation curve T is [a1, a2, ..., an], and the feature vector of the candidate power station curve is [b1, b2, ..., bn]. The similarity metric is selected based on the curve characteristics: cosine similarity is used for stable regions to maintain high accuracy, while Manhattan distance is used for fluctuating regions to improve robustness.

[0088] The method for calculating cosine similarity is as follows:

[0089]

[0090] The Manhattan distance is used to calculate the distance between curves with large fluctuations, and then the similarity is calculated. The calculation method is as follows:

[0091]

[0092] After obtaining the Manhattan distance, the similarity score is calculated as follows:

[0093]

[0094] The similarity between the feature vectors of all photovoltaic power plant sequences in the candidate reference power plant set and the feature vectors of the target power plant sequence T is calculated and quantified into similarity scores. A similarity matrix is ​​then constructed. The similarity calculation results are as follows: Figure 4 As shown.

[0095] Based on the similarity calculation results, two photovoltaic reference power stations with a similarity greater than 0.5 with the target power station T were selected from the candidate power stations: station H (similarity 1.0000) and station I (similarity 0.8857), forming the "dynamic spatial reference photovoltaic station set" for this missing data repair. Weighting coefficients were calculated based on the normalized similarity between each station and T. The weight of station H is calculated. =0.5466, Weight of Station I =0.4534.

[0096] S5. Normalize the similarity scores of each reference power station into weights, and calculate the weighted average of the actual power values ​​of each reference power station at each time point within the missing time period to obtain the preliminary completion value.

[0097] In an optional embodiment of the present invention, step S5 calculates a preliminary complete value by weighting the actual power values ​​of each reference power station at each time point within the missing time period, specifically as follows:

[0098]

[0099] in, This represents the initial completed value for the target photovoltaic power plant at time point t. Let i be the weight of the i-th photovoltaic power station in the dynamic spatial reference power station set. Let be the actual power value of the i-th photovoltaic power station in the dynamic spatial reference power station set at time t.

[0100] In this embodiment, adaptive weight calculation is performed to execute dynamic weighted interpolation in step S5. First, the similarity scores of each reference power station are normalized and converted into weights for each reference power station in the subsequent weighted average. The weight values ​​are positively correlated with the similarity, ensuring that power stations with more similar power generation curves contribute more to the completion result. Then, for each time point t within the missing period, the actual power values ​​of each power station in the dynamic reference set at that time point are weighted and averaged according to their weights to calculate the preliminary completion value of the target power station T.

[0101] In this embodiment, for the missing value of each time point t within the missing time period [t_start, t_end] of the target station T, the interpolated value of the target power station T is calculated by weighting the values ​​of station H and station I with their corresponding time points.

[0102] S6. Calculate the theoretical power generation limit of the target photovoltaic power station by combining meteorological data, apply physical constraints to the preliminary supplementary value, and take the minimum value as the final supplementary value.

[0103] In an optional embodiment of the present invention, step S6 calculates the theoretical power generation limit of the target photovoltaic power station in conjunction with meteorological data, applies physical constraints to the preliminary supplementary value, and takes the minimum value as the final supplementary value, including:

[0104] Using meteorological data from the target power plant and measured irradiance data from meteorological stations for the missing time period, the theoretical upper limit of power generation is calculated through physical modeling. ;

[0105] Apply physical constraints to the initial completion values ​​at each time point to ensure that the final completion values ​​meet the following conditions. .

[0106] This embodiment performs physical law fusion and final output in step S6. Using the component parameters and measured irradiance data of the target power plant, it calculates the theoretical maximum power generation at each time point in the missing period. The preliminary completed value is compared with this theoretical upper limit value, and the smaller one is taken as the final completed data output that conforms to physical laws.

[0107] This embodiment utilizes meteorological factor parameters of the target power plant and measured irradiance data provided by meteorological stations during the missing time period to calculate the theoretical upper limit of power generation through physical modeling. During the interpolation process, the interpolation result at each time step is... Physical constraints are applied to ensure that the interpolation result does not exceed the theoretical upper limit, i.e., the final completed value is: This constraint strictly limits the power generation within the allowable physical feasible range, improving the rationality and reliability of the compensation results. The compensation results for missing values ​​in the time period after physical constraint correction are as follows: Figure 5 As shown.

[0108] Under the same test conditions, Figure 6As shown, the RMSE error between the complemented value and the original data using the method of this invention is 0.16351, while the RMSE errors between the complemented value and the original data using the global mean method, median complemented value method, moving average method, and linear interpolation method are 10.1097, 11.7565, 2.0476, and 2.9464, respectively. Error analysis shows that: (1) The global mean complemented value method and the median complemented value method, which are empirical interpolation methods based on global statistics, are basically ineffective in the task of complementing missing values ​​in the photovoltaic power generation time period. Their errors are significantly higher than other methods, and they cannot effectively reconstruct the time series characteristics. (2) Although the moving average complemented value method and the linear interpolation method can utilize local time series information to improve the complemented value effect to a certain extent, they still have obvious limitations—the moving average method is not responsive enough to rapid fluctuations, and the linear interpolation method is difficult to fit nonlinear changes, resulting in unsatisfactory complemented value accuracy.

[0109] The method proposed in this invention improves the accuracy of missing value imputation by 11.52 times and 17.02 times compared to the moving average imputation method and linear interpolation method, respectively. In summary, the proposed dynamic similarity-weighted imputation method significantly outperforms other traditional missing value imputation methods in terms of accuracy for imputing missing values ​​over a time period. A comparison with the imputation results of traditional methods is shown below. Figure 6 As shown.

[0110] This invention also provides a data time-segmentation system based on deep learning, which applies the interpolation method described above, including:

[0111] The data acquisition module is used to acquire the power generation data and meteorological data of the target photovoltaic power station, identify the time period when the target photovoltaic power station has missing data, and acquire the context time period data of the time period when the data is missing.

[0112] The candidate power station processing module is used to search for photovoltaic power stations within the spatial area where the target photovoltaic power station is located, and to filter and form a set of photovoltaic candidate power stations based on the installed capacity and geographical location information of each photovoltaic power station.

[0113] The temporal feature extraction module is used to extract features from the power generation curves of the photovoltaic candidate power station set and the target photovoltaic power station using an autoencoder based on a temporal convolutional network, and obtain a temporal feature vector.

[0114] The dynamic reference set construction module is used to calculate the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period based on the time-series feature vector, and select the Top-K candidate power stations based on the similarity score to form a dynamic spatial reference power station set;

[0115] The dynamic weighted interpolation module is used to normalize the similarity scores of each reference power station into weights, and to calculate the weighted average of the actual power values ​​of each reference power station at each time point within the missing time period to obtain the preliminary complete value.

[0116] The completion output module is used to calculate the theoretical power generation limit of the target photovoltaic power station by combining meteorological data, apply physical constraints to the preliminary completion value, and take the minimum value as the final completion value.

[0117] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0118] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0120] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0121] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for imputing missing values ​​in data over a time period based on deep learning, characterized in that, Includes the following steps: Acquire the power generation data and meteorological data of the target photovoltaic power station, identify the time period when the target photovoltaic power station has missing data, and obtain the context time period data of the time period when the data is missing; Search for photovoltaic power stations within the spatial area where the target photovoltaic power station is located, and filter them to form a set of candidate photovoltaic power stations based on the installed capacity and geographical location information of each photovoltaic power station; An autoencoder based on a temporal convolutional network is used to extract features from the power generation curves of the photovoltaic candidate power station set and the target photovoltaic power station to obtain a temporal feature vector. The autoencoder based on the temporal convolutional network includes an encoder, a bottleneck layer and a decoder. The encoder extracts temporal features through causal convolution and dilated convolution, the bottleneck layer forces the learning of essential feature representations, and the decoder symmetrically reconstructs the sequence. The encoder includes multiple stacked temporal convolutional network modules, and each temporal convolutional network module uses a causal convolutional layer with a progressive dilation rate to achieve progressive downsampling of the sequence length. The similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period is calculated based on the time series feature vector. The Top-K candidate power stations are selected based on the similarity score to form a dynamic spatial reference power station set. The similarity scores of each reference power station are normalized into weights, and the actual power values ​​of each reference power station are weighted and averaged at each time point within the missing period to obtain the preliminary completion value. Based on meteorological data, the theoretical power generation limit of the target photovoltaic power station is calculated. Physical constraints are applied to the preliminary supplementary value, and the minimum value is taken as the final supplementary value; including: Using meteorological data from the target power plant and measured irradiance data from meteorological stations for the missing time period, the theoretical upper limit of power generation is calculated through physical modeling. ; Apply physical constraints to the initial completion values ​​at each time point to ensure that the final completion values ​​meet the following conditions. ; For each time point within the missing period, the actual power values ​​of each reference power station are weighted and averaged to obtain the preliminary complete value, specifically: in, This represents the initial completed value for the target photovoltaic power plant at time point t. Let i be the weight of the i-th photovoltaic power station in the dynamic spatial reference power station set. Let be the actual power value of the i-th photovoltaic power station in the dynamic spatial reference power station set at time t.

2. The data time-segmentation method based on deep learning according to claim 1, characterized in that, Based on the installed capacity and geographical location information of each photovoltaic power station, a set of candidate photovoltaic power stations was formed, including: The primary criterion is to ensure the consistency of installed capacity, and photovoltaic power plants with capacities different from the target photovoltaic power plant will be eliminated. Based on the time period of missing data for the target photovoltaic power station, photovoltaic power stations whose data is unavailable in the current time period are further excluded.

3. The data time-segmentation method based on deep learning according to claim 1, characterized in that, Progressive downsampling of sequence length is achieved between adjacent modules using convolutional or pooling layers with a stride greater than 1.

4. The data time-segmentation method based on deep learning according to claim 1, characterized in that, The autoencoder based on temporal convolutional networks is trained using a self-supervised learning approach, with the input sequence itself as the reconstruction target, and uses the Adam optimizer to minimize the reconstruction error.

5. The data time-segmentation method based on deep learning according to claim 1, characterized in that, The similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period is calculated based on the time-series feature vector, using at least one of cosine similarity or Manhattan distance, and the measurement method is adaptively selected according to the volatility of the power generation curve.

6. The data time-segmentation method based on deep learning according to claim 5, characterized in that, For the stable region of the power generation curve, cosine similarity is used to calculate the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period; For the fluctuating regions of the power generation curve, the Manhattan distance is used to calculate the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context of the time period.

7. A data time period missing value interpolation system based on deep learning, employing the data time period missing value interpolation method based on deep learning as described in any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to acquire the power generation data and meteorological data of the target photovoltaic power station, identify the time period when the target photovoltaic power station has missing data, and acquire the context time period data of the time period when the data is missing. The candidate power station processing module is used to search for photovoltaic power stations within the spatial area where the target photovoltaic power station is located, and to filter and form a set of photovoltaic candidate power stations based on the installed capacity and geographical location information of each photovoltaic power station. The temporal feature extraction module is used to extract features from the power generation curves of the photovoltaic candidate power station set and the target photovoltaic power station using an autoencoder based on a temporal convolutional network to obtain a temporal feature vector. The autoencoder based on the temporal convolutional network includes an encoder, a bottleneck layer, and a decoder. The encoder extracts temporal features through causal convolution and dilated convolution, the bottleneck layer forces the learning of essential feature representations, and the decoder symmetrically reconstructs the sequence. The encoder includes multiple stacked temporal convolutional network modules, and each temporal convolutional network module uses a causal convolutional layer with a progressive dilation rate to achieve progressive downsampling of the sequence length. The dynamic reference set construction module is used to calculate the similarity score between the target photovoltaic power station and the photovoltaic candidate power station in the context time period based on the temporal feature vector, and select the Top-K candidate power stations based on the similarity score to form a dynamic spatial reference power station set; The dynamic weighted interpolation module is used to normalize the similarity scores of each reference power station into weights, and to calculate the weighted average of the actual power values ​​of each reference power station at each time point within the missing time period to obtain the preliminary complete value. The completion output module is used to calculate the theoretical power generation limit of the target photovoltaic power station by combining meteorological data, apply physical constraints to the preliminary completion value, and take the minimum value as the final completion value.

Images