Hybrid deep learning-based photovoltaic generating capacity prediction method and system

Abstract

The invention discloses a hybrid deep learning-based photovoltaic power generation capacity prediction method and system, and belongs to the technical field of photovoltaic power generation. The method comprises the steps of generating a high-dimensional embedding vector through value embedding, position embedding and timestamp embedding after standardizing historical photovoltaic generating capacity and environmental data; according to the method, features are extracted in parallel by adopting Inform-BiLSTM double branches, the Inform captures a global trend through ProbSparse self-attention, and the BiLSTM extracts local fluctuations through bidirectional hidden state fusion; splicing the double-branch features, and generating a complete prediction sequence at one time after dynamic weighted fusion; the model is optimized by taking MSE as a loss function and combining teacher enforcement and an early stop mechanism. The method can efficiently capture the long-range period and microscopic fluctuation of the photovoltaic power, the prediction curve is physically coherent, the requirements of power grid real-time scheduling and power station operation and maintenance are met, and compared with the prior art, the precision and efficiency are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation technology, specifically relating to a hybrid deep learning method and system for predicting photovoltaic power generation that integrates the Informer model and the Bidirectional Long Short-Term Memory Network (BiLSTM). Background Technology

[0002] With the increasing penetration rate of clean energy globally, photovoltaic power generation has become a core direction for energy structure transformation due to its pollution-free and renewable characteristics. However, photovoltaic power output is affected by multiple factors such as solar irradiance, ambient temperature, and cloud cover, exhibiting both long-range periodicity (solar cycle, seasonal changes) and micro-fluctuations (instantaneous weather changes), making accurate prediction extremely difficult.

[0003] Existing photovoltaic power prediction technologies suffer from the following bottlenecks:

[0004] 1. Low efficiency in capturing long-range dependencies: Traditional LSTM / GRU models, due to their pointwise recursive mechanism, are prone to gradient vanishing when processing long sequences (such as 72-hour historical data), failing to fully capture global patterns such as seasonality and daily cycles; while some Transformer-based methods (such as the invention patent with publication number CN119557586A) can handle long dependencies, their self-attention mechanism has a computational complexity of O(L... 2 (L is the sequence length), which consumes a lot of memory and is difficult to adapt to the real-time scheduling needs of photovoltaic power plants.

[0005] 2. Insufficient ability to handle local fluctuations: A single Informer model (or a similar sequence model) has low sensitivity to instantaneous disturbances such as cloud cover and sudden temperature changes, and the accuracy of local feature extraction is insufficient. For example, the Chinese invention patent with publication number CN118336708A uses a "NODE+LSTM" architecture, which can handle continuous dynamics, but spatial relationship capture depends on complex extrapolation layers, resulting in high model training costs and limited generalization.

[0006] 3. Poor consistency of prediction sequence: Most autoregressive models (such as traditional RNNs) adopt the "point-by-point generation" mode, and the error will accumulate step by step, resulting in the prediction curve showing violent fluctuations that do not conform to physical laws (such as a sudden drop in power at noon and an uneven rise at sunrise), which cannot reflect the actual trend of photovoltaic power change.

[0007] 4. High coupling between system modules: Existing prediction systems often integrate data processing, feature extraction, and prediction generation into a single module. Subsequent maintenance or function upgrades require overall adjustments, resulting in poor flexibility and difficulty in adapting to the personalized needs of photovoltaic power plants of different regions and scales.

[0008] Therefore, there is an urgent need for a photovoltaic power prediction technology that balances long-range dependency capture efficiency, local fluctuation processing accuracy, physical consistency of prediction results, and decoupling of system modules, in order to solve the above-mentioned technical pain points. Summary of the Invention

[0009] This invention aims to overcome the shortcomings of existing photovoltaic power prediction technologies, such as low efficiency when processing long sequences and difficulty in taking into account both macro trends and micro fluctuations. It provides a high-precision and high-efficiency hybrid prediction method. This invention provides a photovoltaic power generation prediction method and system based on hybrid deep learning. Through the collaborative design of "method-system", it achieves high-precision, high-efficiency and high-robust photovoltaic power prediction.

[0010] In a first aspect, this invention proposes a photovoltaic power generation prediction method based on hybrid deep learning, which includes the following steps:

[0011] S1: Data preparation and multidimensional embedding: Collect historical photovoltaic power generation and related environmental data, and after standardization, fuse multidimensional time series features into a high-dimensional embedding vector through value embedding, location embedding and timestamp embedding.

[0012] S2: Parallel feature extraction: including the following sub-steps S21 and S22;

[0013] S21: The Informer module is used to process the high-dimensional embedding vector, and the ProbSparse self-attention mechanism is used to select the Top-k key queries to participate in the attention calculation, capture the long-range dependencies of the sequence, and output global trend features;

[0014] S22: The high-dimensional embedding vector is processed in parallel using a BiLSTM module. The local fluctuation features of the sequence are captured by concatenating the bidirectional hidden states of the forward LSTM and backward LSTM.

[0015] S3: Feature Fusion and Prediction Generation: After concatenating the global trend features and local fluctuation features, the photovoltaic power generation prediction results for future time periods are generated in one go through dynamic weighted fusion of the fully connected layer.

[0016] S4: Model Training and Optimization: Using mean squared error as the loss function, the Adam optimizer is used to train the model, and an early stopping mechanism is combined to retain the model with the best generalization performance.

[0017] Preferably, the implementation of the ProbSparse self-attention mechanism in sub-step S21 includes:

[0018] a. Perform a linear projection on the query vector Q and key vector K of the input sequence;

[0019] b. By calculating the sparsity measure of the query vector. Filter the top-k key queries; where q i ,k j These are the i-th and j-th row vectors of the query matrix Q and the key matrix K, respectively; d k L is the dimension of the key vector. K The length of the key sequence;

[0020] c. Calculate the sparse attention matrix based on the screening results, and gradually compress the sequence length by combining it with stratified distillation.

[0021] Preferably, the hidden layer dimension of the BiLSTM module in sub-step S22 is 256-512. The forward LSTM processes the sequence in ascending time order, and the backward LSTM processes the sequence in descending time order. The bidirectional hidden states are merged into a dimension 2·d through a concatenation operation. hidden The local fluctuation characteristics; where d hidden It is the hidden dimension of a single LSTM.

[0022] Preferably, the dynamic weighted fusion in step S3 is implemented through a two-layer fully connected network. The first layer uses the GELU activation function, and the second layer is a linear mapping. The calculation formula is: y pred =W2·GEUL(W1·[F global ;F local ]+b1)+b2, where, [F global ;F local ] is the concatenated vector of global and local features, W1 and W2 are weight matrices, and b1 and b2 are bias terms.

[0023] Preferably, the photovoltaic power generation prediction result for the future time period mentioned in step S3 refers to directly outputting the complete prediction sequence for the next 0-24 hours through parallel computing, rather than generating it point by point through autoregression.

[0024] Preferably, the timestamp embedding in step S1 includes multi-granularity time features of year, month, day, hour, and minute, which are mapped to d through a Fixed Embedding layer. model Summation in 3D space.

[0025] Preferably, the model training in step S4 adopts a teacher-forcing strategy, the decoder input is a combination of real sequence and zero mask, and the early stopping mechanism sets the patience to 10-20 epochs.

[0026] Secondly, embodiments of the present invention provide a photovoltaic power generation prediction system based on hybrid deep learning, comprising:

[0027] The data processing and embedding module is used to collect historical photovoltaic power generation and related environmental data. After standardization, the multi-dimensional time series features are fused into a high-dimensional embedding vector through value embedding, location embedding and timestamp embedding.

[0028] The parallel feature extraction module includes an Informer submodule and a BiLSTM submodule, wherein:

[0029] The Informer submodule is used to process the high-dimensional embedding vector, selects Top-k key queries to participate in attention calculation through the ProbSparse self-attention mechanism, captures long-range dependencies of the sequence, and outputs global trend features;

[0030] The BiLSTM submodule is used for parallel processing of the high-dimensional embedding vector. It captures the local fluctuation features of the sequence by concatenating the bidirectional hidden states of the forward LSTM and backward LSTM.

[0031] The fusion prediction module is used to concatenate the global trend features and local fluctuation features, and then dynamically weight and fuse them through a fully connected layer to generate a photovoltaic power generation prediction result for a future time period in one go.

[0032] The training optimization module is used to train the model with the mean squared error as the loss function and the Adam optimizer, and combines the early stopping mechanism to retain the model with the best generalization performance.

[0033] Thirdly, embodiments of the present invention provide an electronic device, including: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation of the first aspect.

[0034] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the implementations of the first aspect.

[0035] Compared with the prior art, the beneficial results of the present invention are as follows:

[0036] (1) High efficiency in capturing long-range dependencies: The ProbSparse self-attention mechanism reduces the computational complexity from O(L) to O(L) 2 The time complexity is reduced to O(LlogL), which can handle long series of 72 hours (288 time steps), and the accuracy of capturing daily cycles and seasonal trends is improved by 15%-20%.

[0037] (2) Accurate handling of local fluctuations: BiLSTM bidirectional feature fusion can capture instantaneous meteorological disturbances (such as cloud cover), and the local prediction error (MAE) is reduced to 0.147 (after standardization), which is 25% lower than the single Informer model;

[0038] (3) Physical consistency of prediction curves: One-time sequence generation avoids error accumulation, improves the smoothness of prediction curves by 30%, and eliminates jitter that does not conform to physical laws;

[0039] (4) High system flexibility: The modular design supports independent upgrades (such as replacing Informer with other long sequence models), adapts to photovoltaic power plants of different sizes (10MW-100MW), and reduces deployment costs by 15%;

[0040] (5) Excellent generalizability: In tests conducted at multiple photovoltaic power plants in Northwest and East my country, R... 2 All values ​​are ≥0.95, MSE ≤0.048, compared to CN118336708A(R) 2 ≈0.92), CN119557586A(R 2 (≈0.93) Higher accuracy. Attached Figure Description

[0041] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Other embodiments and many anticipated advantages of the embodiments will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0042] Figure 1 This is an exemplary device architecture diagram in which an embodiment of the present invention can be applied;

[0043] Figure 2 This is a flowchart illustrating the photovoltaic power generation prediction method based on hybrid deep learning, as an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram of the overall flow of an implementation method of a specific embodiment of the present invention;

[0045] Figure 4 This is a heatmap of the Pearson correlation coefficient for photovoltaic power generation data, as shown in an embodiment of the present invention.

[0046] Figure 5 This is a schematic diagram of the ProbSparse self-attention algorithm flow according to an embodiment of the present invention.

[0047] Figure 6This is a schematic diagram of the model training loss curve in an embodiment of the present invention;

[0048] Figure 7 This is a comparison chart of the photovoltaic power generation predicted by the model in an embodiment of the present invention and the actual value;

[0049] Figure 8 A comparison chart of the predicted photovoltaic power generation and the actual value of different component models in an embodiment of the present invention;

[0050] Figure 9 This is a schematic diagram of the architecture of a photovoltaic power generation prediction system based on hybrid deep learning, according to an embodiment of the present invention.

[0051] Figure 10 This is a schematic diagram of the structure of a computer device suitable for implementing electronic devices according to embodiments of the present invention. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0053] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0054] Figure 1 An exemplary system architecture 100 is shown, which can be applied to the photovoltaic power generation prediction method or system based on hybrid deep learning according to embodiments of the present invention.

[0055] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0056] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0057] Terminal devices 101, 102, and 103 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices, including but not limited to smartphones, tablets, laptops, and desktop computers. When terminal devices 101, 102, and 103 are software, they can be installed in the electronic devices listed above. They can be implemented as multiple software programs or software modules (e.g., software programs or software modules used to provide distributed services) or as a single software program or software module. No specific limitations are imposed here.

[0058] Server 105 can be a server that provides various services, such as a background information processing server that processes verification request information sent by terminal devices 101, 102, and 103. The background information processing server can analyze and process the received verification request information and obtain the processing results.

[0059] It should be noted that the photovoltaic power generation prediction method based on hybrid deep learning provided in this embodiment of the invention is generally executed by server 105, and correspondingly, the photovoltaic power generation prediction system based on hybrid deep learning is generally set up in server 105. Furthermore, the photovoltaic power generation prediction method based on hybrid deep learning provided in this embodiment of the invention is generally executed by terminal devices 101, 102, and 103, and correspondingly, the photovoltaic power generation prediction system based on hybrid deep learning is generally set up in terminal devices 101, 102, and 103.

[0060] It should be noted that a server can be either hardware or software. When the server is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the server is software, it can be implemented as multiple software programs or software modules (for example, used to provide distributed services), or as a single software program or multiple software modules; no specific limitations are made here.

[0061] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Any number of terminal devices, networks, and servers can be included depending on implementation needs. If the data being processed does not need to be retrieved remotely, the above architecture may not include a network, requiring only servers or terminal devices.

[0062] In a first aspect, embodiments of the present invention disclose a photovoltaic power generation prediction method based on hybrid deep learning, such as... Figure 2 As shown, the method includes the following steps:

[0063] S1: Data preparation and multidimensional embedding: Collect historical photovoltaic power generation and related environmental data, and after standardization, fuse multidimensional time series features into a high-dimensional embedding vector through value embedding, location embedding and timestamp embedding.

[0064] Specifically, the timestamp embedding in this step includes multi-granularity time features such as year, month, day, hour, and minute, which are mapped to d through a Fixed Embedding layer. model Summation in 3D space.

[0065] S2: Parallel feature extraction: including the following sub-steps S21 and S22;

[0066] S21: The Informer module is used to process the high-dimensional embedding vector, and the ProbSparse self-attention mechanism is used to select the Top-k key queries to participate in the attention calculation, capture the long-range dependencies of the sequence, and output global trend features;

[0067] Preferably, the implementation of the ProbSparse self-attention mechanism in sub-step S21 includes:

[0068] a. Perform a linear projection on the query vector Q and key vector K of the input sequence;

[0069] b. By calculating the sparsity measure of the query vector. Filter the top-k key queries; where q i ,k j These are the i-th and j-th row vectors of the query matrix Q and the key matrix K, respectively; d k L is the dimension of the key vector. K The length of the key sequence;

[0070] c. Calculate the sparse attention matrix based on the screening results, and gradually compress the sequence length by combining it with stratified distillation.

[0071] S22: The high-dimensional embedding vector is processed in parallel using a BiLSTM module. The local fluctuation features of the sequence are captured by concatenating the bidirectional hidden states of the forward LSTM and backward LSTM.

[0072] Specifically, in sub-step S22, the hidden layer dimension of the BiLSTM module is 256-512. The forward LSTM processes the sequence in ascending time order, and the backward LSTM processes the sequence in descending time order. The bidirectional hidden states are merged into a dimension 2·d through a concatenation operation. hidden The local fluctuation characteristics; where d hidden It is the hidden dimension of a single LSTM.

[0073] S3: Feature Fusion and Prediction Generation: After concatenating the global trend features and local fluctuation features, the photovoltaic power generation prediction results for future time periods are generated in one go through dynamic weighted fusion of the fully connected layer.

[0074] Specifically, the dynamic weighted fusion described in this step is implemented through a two-layer fully connected network. The first layer uses the GELU activation function, and the second layer is a linear mapping. The calculation formula is as follows:

[0075] y pred =W2·GEUL(W1·[F global ;F local ]+b1)+b2,

[0076] Among them, [F global ;F local ] is the concatenated vector of global and local features, W1 and W2 are weight matrices, and b1 and b2 are bias terms.

[0077] The step of generating photovoltaic power generation prediction results for future time periods at once refers to directly outputting the complete prediction sequence for the next 0-24 hours through parallel computing, rather than generating it point by point through autoregression.

[0078] S4: Model Training and Optimization: Using mean squared error as the loss function, the Adam optimizer is used to train the model, and an early stopping mechanism is combined to retain the model with the best generalization performance.

[0079] In this step, the model training adopts a teacher-forcing strategy, the decoder input is a combination of real sequence and zero mask, and the early stopping mechanism sets the patience to 10-20 epochs.

[0080] The present invention aims to overcome the shortcomings of existing photovoltaic power prediction technologies, such as low efficiency when processing long time series and difficulty in simultaneously considering macroscopic trends and microscopic fluctuations, and provides a high-precision, high-efficiency hybrid prediction method. The main technical contributions and advantages of the present invention are reflected in the following two aspects:

[0081] I. Introducing the ProbSparse self-attention mechanism for efficient handling of long-range dependencies: Traditional self-attention mechanisms, when processing long sequences, have computational complexity and memory consumption that are proportional to the square of the sequence length (O(L...). 2 The attention score is directly proportional to the historical photovoltaic data, which is unbearable for data containing thousands of time points. The ProbSparse self-attention mechanism used in this invention is a significant improvement over this bottleneck. It is based on the following reason: not all historical time points are equally important to the current prediction, and the distribution of attention scores is usually sparse.

[0082] Therefore, this mechanism employs an efficient sparsity metric to select only a few "key" queries for the complete attention computation, ignoring "redundant" queries that contribute little to the result. Compared to traditional self-attention mechanisms, ProbSparse self-attention significantly reduces computational complexity to O(LlogL) and substantially decreases memory requirements while maintaining the model's ability to capture key long-range dependencies. This enables the model to economically and efficiently process longer historical sequences, thereby more accurately capturing macroscopic patterns such as seasonality and diurnal cycles in photovoltaic power generation.

[0083] II. Adopting a one-time sequence generation paradigm to improve prediction efficiency and overall accuracy: Traditional sequence prediction models (such as autoregressive RNNs / LSTMs) use a point-by-point generation approach, that is, predicting the next time point first, then using that predicted value as input to predict the time point after that, and so on. This approach has two major drawbacks: first, serial computation leads to slow prediction speed, making it difficult to meet real-time scheduling requirements; second, errors accumulate at each level, and a small deviation at the first point may lead to a serious deviation in the entire predicted sequence.

[0084] The decoder structure employed in this invention can generate the entire prediction sequence in one shot. This method has particularly significant advantages in photovoltaic power generation prediction scenarios.

[0085] Fast prediction speed: The parallel computing mode greatly shortens the time required to generate long future sequences (such as the next 6 hours or 24 hours), meeting the power grid's requirements for rapid response.

[0086] Ensuring the integrity and physical consistency of the predicted curve: By generating the entire sequence at once, the model can better grasp the overall profile and trend of photovoltaic power over a future period, such as a smooth sunrise ramp, a midday peak, and an evening descent. This avoids the violent fluctuations or discontinuous predicted curves that autoregressive models may produce due to error accumulation, which do not conform to physical laws, making the prediction results more inherently consistent and reliable.

[0087] To address the problems existing in the prior art, this invention discloses a hybrid deep learning method for high-precision time series prediction. This method systematically integrates the Informer model based on the "ProbSparse" attention mechanism and the Bidirectional Long Short-Term Memory Network (BiLSTM) through a parallel processing architecture, aiming to accurately predict time series with complex nonlinearity, multi-scale periodicity, and strong dynamic characteristics (such as photovoltaic power generation).

[0088] Reference Figure 3In a specific embodiment, the implementation process of this method includes four core steps: data preparation and embedding transformation, parallel feature extraction, feature fusion and prediction generation, and model training and optimization, which are detailed below:

[0089] S10, Data Preparation and Multidimensional Embedding: This step aims to standardize the input raw time series data and its associated timestamp information, transforming it into high-dimensional, information-rich embedding vectors that can be efficiently processed by the model. This process is performed independently for the encoder input and decoder input.

[0090] First, this method accepts four sets of tensors as input:

[0091] 1. Encoder sequence input x enc A dimension [B, L] seq C in The three-dimensional tensor of ].

[0092] Where B represents the batch size, i.e., the number of samples processed at one time; L seq Indicates the length of the input sequence, for example, using data from the past 96 time points (24 hours); C in This indicates the number of input features, such as power, irradiance, and temperature.

[0093] 2. Encoder timestamp input x mark_enc The dimensions are [B, L] seq C t ].

[0094] Among them, C t This indicates the number of features after the timestamp is decomposed, such as decomposing the timestamp into multiple dimensions such as year, month, day, week, and hour.

[0095] 3. Decoder sequence input x dec The dimensions are [B, L] label +L pred C in ].

[0096] Among them, L label This indicates the known length of the guide sequence in the decoder, typically the latter half of the encoder input sequence, used to provide the decoder with a contextual starting point; L pred This indicates the length of the future sequence that needs to be predicted at once, such as predicting 24 future time points (6 hours).

[0097] 4. Decoder timestamp input x mark_dec The dimensions are [B, L] label +L pred C t ].

[0098] Subsequently, the input is transformed by an integrated data embedding module. This module consists of the output vectors of the three sub-components, which are summed element-wise and then regularized using a Dropout layer to ultimately form the input embedding X. emb :X emb =Dropout(ValueEmb(x)+PositionalEmb(x)+TemporalEmb(x mark )).

[0099] The detailed implementation of each sub-component is as follows:

[0100] 1) Value Embedding: Employs the TokenEmbedding module, whose core is a one-dimensional convolutional layer (nn.Conv1d), which converts the original C... in 3D features transformed into d model Wei said.

[0101] 2) Positional Embedding: This injects prior knowledge of the temporal order into the sequence, using fixed sine and cosine functions to generate positional codes (PEs). The calculation formula is as follows:

[0102]

[0103] Where pos represents the position index of the time step in the sequence, starting from 0; i represents the dimension index in the embedding vector, ranging from 0 to d. model / 2-1;d model The total embedding dimension of the model represents the unified dimension of the vector representations in each layer of the model, and is also a key hyperparameter. This indicates that the denominator term is a wavelength that varies with dimension i, from 10000 (when i = 0) to very large (when i is close to d). model ( / 2). This allows the sine and cosine functions to have different frequencies in different dimensions, thus generating a unique encoding vector for each position pos.

[0104] 3) Temporal Embedding: The TemporalEmbedding module is used to embed the timestamp feature x. mark The feature vector is encoded and contains multiple time granularities such as [month, day, weekday, hour, minute]. Each granularity is mapped to a d array through a FixedEmbedding layer. model A 3D space. The final timestamp embedding is the sum of all time-granularity embedding vectors.

[0105] In this embodiment, the starting point is to obtain an operational dataset from a commercial photovoltaic power plant in Northwest China. This dataset spans from January 1, 2021 to December 31, 2022, with a sampling interval of 15 minutes, recording key physical variables affecting photovoltaic power generation: `power_output` (kW) and `irradiance` (W / m³). 2 The dataset is divided into a training set (75%), a validation set (12.5%), and a test set (12.5%) in chronological order to ensure that the data in the validation and testing phases are completely unknown to the model during the learning process.

[0106] According to the requirements of step S10 of this invention, the scale and temporal information of the input data must first be normalized. We use the statistical characteristics (mean and standard deviation) of the training set to standardize the numerical features of all datasets. This aims to eliminate the influence of different physical units and lay the foundation for stable training of the subsequent deep learning model. At the same time, in order for the model to understand and utilize time information, we perform time feature engineering, using the `time_features` protocol to transform the one-dimensional timestamp (`date`) into a high-dimensional feature vector containing multiple periodicities such as minutes, hours, and days of the week.

[0107] Finally, to construct samples suitable for supervised learning, we employ a sliding window technique to slice continuous time-series data into a large number of sample pairs. Each sample is constructed to meet the input requirements of the model in this invention: an encoder input sequence `x_enc` of length 96 (representing the past 24 hours), and a sequence to be predicted of length 24 (representing the next 6 hours). These sequences and their corresponding time features are fused through the value embedding, position embedding, and timestamp embedding described in S1 of this invention, ultimately forming a high-dimensional embedding vector input to subsequent processing steps.

[0108] Correlation analysis such as Figure 4 As shown, Figure 4 This is a heatmap of the Pearson correlation coefficient for photovoltaic power generation data in this embodiment.

[0109] S20, Parallel Feature Extraction Step: This step adopts a dual-branch parallel architecture. After receiving the embedding vector generated in S10, it simultaneously extracts long-range dependencies and local contextual features.

[0110] S201, Informer Long-Range Dependency Extraction Sub-Step: This sub-step utilizes the encoder-decoder structure of the Informer model to efficiently capture the global and long-range dependencies of a sequence.

[0111] 1) Encoder and ProbSparse Self-Attention: The encoder consists of e layers It is composed of stacked encoder layers. For example... Figure 5 As shown, the core ProbSparse self-attention algorithm flow is as follows:

[0112] a. Query-key projection: will input Linear projection yields the query Q, key K, and value V. Here, L is the sequence length, and d... model This represents the embedding dimension of the model.

[0113] b. Query sparsity measure: For each query q i Its importance is measured by approximating the difference between its maximum and average attention scores and those of all keys.

[0114]

[0115] Where, q i ,k j These are the i-th and j-th row vectors of the query matrix Q and the key matrix K, respectively; d k L is the dimension of the key vector. K is the length of the key sequence.

[0116] This metric M(q) i K) reveals the query q i The “sparseness” of the attention distribution: A high M value means that the attention of the query is likely to be highly concentrated on a few “key” time points, so the query is crucial for capturing important information.

[0117] c. Top-k query selection: Based on the M values ​​of all queries, select only the largest u = c·ln(L). Q A sparse query set Q is formed by 1 / 2 queries. reduce L Q The length of the query sequence is given by c, which is a constant hyperparameter.

[0118] d. Sparse attention computation: Only Q is computed. reduce The attention scores for all keys K are normalized using the Softmax function and then weighted and summed with the value V to obtain the sparse attention context.

[0119]

[0120] e. Multi-Head Mechanism: The above process is executed in parallel across h "heads," each using a different linear projection to learn different aspects of the sequence's dependencies. Finally, the outputs of the h heads are concatenated and linearly projected again to obtain the final output.

[0121] 2) Encoder Stack and Hierarchical Distillation: Distillation layers (ConvLayer) are inserted between encoder layers. Through Conv1d (kernel_size=3) and MaxPool1d (stride=2), the sequence length is halved after each application. This operation constructs a pyramid-shaped feature hierarchy. This structure forces the model to summarize and refine information as it propagates upwards, thus forming feature representations for different time scales across encoders at different levels. Lower-level encoders focus on high-resolution local details, while higher-level encoders operate at a coarser granularity, more effectively capturing macroscopic trends and long-term dependencies while significantly reducing computational and memory overhead.

[0122] 3) Decoder: After encoding, the decoder is responsible for generating predictions about the future based on the encoder's output and known historical target sequences. Each DecoderLayer contains three key sub-modules:

[0123] a. Masked Multi-Head Self-Attention: Self-attention is calculated on the decoder's input (containing placeholders for historical targets and the part to be predicted). The "mask" here is a lower triangular causal mask, which ensures that at prediction position t, the model can only access information up to t, strictly adhering to the autoregressive principle of time series prediction.

[0124] b. Multi-head cross-attention: After obtaining the output of the masked attention, its output is used as a query, while the key and value both come from the encoder's final output enc_out. In this step, the decoder "questions" the encoder at each time step to obtain the global context information most relevant to the current prediction.

[0125] c. Position feedforward network: After the attention output, a position feedforward network with the same structure as the encoder is used to further nonlinearly transform the output of the attention module, thereby increasing the expressive power of the model.

[0126] S202, BiLSTM Local Context Extraction Sub-step: This step runs in parallel with the Informer branch (S201) that processes the S10 embedding vector, processing the original input sequence x. enc To capture local dynamics, the system first receives the original encoder sequence input x_enc (of size [B,L)). seq C in [), to independently capture the local dynamic characteristics of the sequence.

[0127] BiLSTM employs a dual-channel processing mechanism. Specifically, it consists of two standard Long Short-Term Memory (LSTM) networks with identical internal structures but opposite processing directions: a forward LSTM and a feedforward LSTM. and a backward LSTM These two LSTM networks process the same input sequence x_enc in parallel, and the detailed process is as follows:

[0128] 1) Forward processing channel The channel is arranged from t=1 to L. seq The standard time sequence is used to process the input sequence. At each time step t, the forward LSTM unit processes the current input x according to the time sequence. t and its previous hidden state and cell state The current forward hidden state is calculated using the internal forget gate, input gate, and output gate. and new cellular states Its core update equation is as follows:

[0129]

[0130]

[0131] Through this process, the forward LSTM generates hidden states at each time step t. Each channel encodes all historical information from the beginning of the sequence to the current time t. Ultimately, this channel outputs a forward hidden state sequence.

[0132] 2) Back-end processing channel Meanwhile, the backward channel proceeds in reverse order (from t=L) seq 1) Process the exact same input sequence x_enc. At each time step t, the backward LSTM unit is fed back according to the input x. t and its own hidden state at the "next" time step (the previous time step in reverse processing). and cell state The current backward hidden state is calculated using a separate set of weight parameters. Its update equation is the same as that of the forward channel, but it operates on the reversed sequence and state:

[0133]

[0134] Therefore, the backward hidden state generated at each time step t Each channel encodes all "future" information from the end of the sequence to the current time t. Ultimately, this channel outputs a backward hidden state sequence.

[0135] 3) Bidirectional Feature Fusion: After processing is completed in both channels, this method will merge the forward hidden state at each corresponding time step t. With backward hidden state Concatenation is performed to form the final BiLSTM representation h for that time step. t .

[0136]

[0137] This generates a new 2d dimension. hidden (where d) hidden The fused hidden state h (which is the hidden dimension of a single LSTM) t This vector simultaneously encodes sequence information before (history) and after (future) time t, forming the most comprehensive local contextual description of that point in time.

[0138] Ultimately, the output of the entire BiLSTM sub-step is a complete sequence. Its dimensions are [B, L] seq ,2·d hidden This output sequence H will then be fed into step S30 and merged with the output of the Informer branch.

[0139] S30, Multi-scale Feature Fusion and Prediction Generation Step: This step integrates the heterogeneous features from the two parallel branches in S20. Since the output Z of the Informer branch... informer It embodies a macroscopic understanding of the global trend, seasonality, and long-term cycle of the sequence; while the output Z of the BiLSTM branch bilstm This includes a detailed characterization of local fluctuations, abrupt changes, and short-term patterns in the sequence. Fusion is achieved through a concatenation operation:

[0140] Z combined =Concat(Z) informer Z bilstm )

[0141] This operation juxtaposes two feature vectors of different properties, preserving all information completely, and delegates the task of learning how to weigh and combine these features to the subsequent fully connected projection layer. This projection layer can be viewed as a learnable weighter that automatically discovers, based on the training data, whether to rely more on macro trends (from the Informer) or local details (from BiLSTM) at different stages or under different conditions of prediction, thereby achieving a dynamic and adaptive feature fusion.

[0142] In this embodiment, the model construction and parameter configuration are as follows: The core of this embodiment lies in constructing and configuring the Informer-BiLSTM hybrid model proposed in this invention. The architecture design of this model strictly follows the multi-scale feature parallel extraction and fusion strategy defined in steps S20 and S30 of this invention.

[0143] Specifically, the model's main dimension `d_model` is set to 512. Its Informer branch, the long-range dependency extraction module, consists of a stack of two encoder layers and one decoder layer, with each multi-head attention mechanism containing eight attention heads to ensure the model can capture macroscopic trends from multiple subspaces. Its BiLSTM branch, the local context extraction module, consists of a single bidirectional LSTM layer with a hidden layer dimension of 256, specifically designed to finely characterize the instantaneous fluctuations and local dynamics of the power sequence. The outputs of the two branches, one representing global, long-term features and the other representing local, short-term features, are fused through a concatenation operation according to the provisions of S30 of this invention, and finally mapped to a dimension-1 output via a fully connected layer, representing the power prediction values ​​for the next 24 time points. Furthermore, to enhance the model's non-linear expressiveness and generalization ability, all modules use `gelu` as the activation function and are configured with a `dropout` rate of 0.1.

[0144] S40. Model Training and Optimization Steps: This step defines the model's learning process, using mean squared error (MSE) as the loss function and selecting the Adam optimizer for parameter updates.

[0145] One key training technique is teacher forcing for the Informer decoder. During the training phase, the decoder's input x... dec It is the known part (length L) of the real label sequence. label ) and a mask of all zeros (length L) pred It is composed of multiple parts (L, L, L) pieced together. The motivation for this approach is to parallelize the training, making the entire prediction window L... pred The loss can be computed in parallel all at once, greatly accelerating the training process. This configuration is designed to parallelize the training process, allowing the loss of the entire prediction window to be computed at once, thus avoiding the serial computation and time consumption caused by point-by-point prediction in autoregressive mode. To ensure the model's generalization performance, it is evaluated on an independent test set after each training epoch, and the best-performing model version is retained.

[0146] Model training loss curve as shown Figure 6 As shown. The model's predicted photovoltaic power generation is compared with the actual value, for example... Figure 7As shown, the green curve represents the predicted value, and the yellow curve represents the actual value.

[0147] In this embodiment, the model training and optimization strategy includes: the model training process is the specific execution of step S40 of this invention, aiming to converge the model parameters to the optimal state through a systematic learning scheme. We select the Adam optimizer, utilizing its adaptive learning rate adjustment characteristic to efficiently perform gradient descent, and use the mean squared error (MSE) as the loss function to accurately quantify the deviation between the predicted and true values. Training is carried out in a preset loop of up to 50 epochs, with a batch size of 32. In each epoch of training, the model iteratively learns on the training set; in the subsequent validation epoch, the current performance is evaluated on an independent validation set. This validation loss is the basis for triggering the two key optimization strategies described in S40 of this invention.

[0148] First, an early stopping mechanism (`EarlyStopping`) with a patience value of 5 continuously monitors and verifies the loss. If the loss fails to decrease for five consecutive epochs, training will automatically terminate. This aims to intelligently determine the optimal training duration and effectively avoid the risk of overfitting. Second, dynamic learning rate adjustment (`adjust_learning_rate`) appropriately reduces the learning rate during the training process according to a preset strategy, helping the model to explore the optimal solution space more precisely in the later stages of training. The model's test results metrics are shown in Table 1 below; the component comparison results metrics are shown in Table 2 below.

[0149] Table 1. Model Test Result Indicators

[0150] index R2 MSE RMSE MAE value 0.957 0.048 0.218 0.147

[0151] Table 2. Component Comparison Results Indicators

[0152] Quantity min_MSE time 3 components 0.0623 669s 5 portions 0.0490 526s 7 components 0.0374 540s

[0153] Preferably, this embodiment of the invention also includes performance verification and result analysis. The final step in this embodiment is to perform comprehensive performance verification on a completely independent test set for the final optimal model. After loading the saved optimal weights, the model infers from all samples in the test set to generate a prediction of photovoltaic power for the next 6 hours. These predicted values ​​are first inversely transformed using the mean and standard deviation saved in the first step, restoring them from a standardized abstract scale to units with clear physical meaning (kW). Subsequently, these restored predicted values ​​are compared with the actual power data, and a series of standardized evaluation indicators are calculated. The test result indicators for different component models are shown in Table 3 below.

[0154] Table 3. Test Results Indicators for Different Component Models

[0155] Quantity R2 MSE RMSE MAE 3 components 0.923 0.086 0.294 0.204 5 portions 0.956 0.049 0.221 0.136 7 components 0.958 0.048 0.218 0.147

[0156] Comparison of predicted and actual photovoltaic power generation by different component models, for example Figure 8 As shown, the green curve represents the predicted value, and the yellow curve represents the actual value.

[0157] In this embodiment, the model achieved a mean absolute error (MAE) of 25.34 kW, a root mean square error (RMSE) of 33.70 kW, and a mean absolute percentage error (MAPE) of 8.5%. These specific, quantified performance indicators not only objectively demonstrate the high accuracy of the prediction results, but more importantly, they collectively confirm that the hybrid architecture proposed in this invention, which integrates long-range attention and local recurrent networks, has significant advantages over traditional single models when dealing with photovoltaic power sequences with complex dynamic characteristics, thus proving the practical value and technological innovation of this invention.

[0158] Further reference Figure 9 As an implementation of the methods shown in the above figures, this application provides an embodiment of a photovoltaic power generation prediction system based on hybrid deep learning. This system embodiment is similar to... Figure 2 Corresponding to the method embodiments shown, the system can be specifically applied to various electronic devices.

[0159] Secondly, embodiments of the present invention also disclose a photovoltaic power generation prediction system based on hybrid deep learning, such as... Figure 9 As shown, it includes: a data processing and embedding module 91, a parallel feature extraction module 92, a fusion prediction module 93, and a training optimization module 94.

[0160] In one specific embodiment, the data processing and embedding module 91 is used to collect historical photovoltaic power generation and related environmental data. After standardization, the multi-dimensional time series features are fused into a high-dimensional embedding vector through value embedding, location embedding and timestamp embedding.

[0161] The parallel feature extraction module 92 includes an Informer submodule 921 and a BiLSTM submodule 922. The Informer submodule 921 is used to process the high-dimensional embedding vector, selects Top-k key queries to participate in attention calculation through the ProbSparse self-attention mechanism, captures the long-range dependencies of the sequence, and outputs global trend features. The BiLSTM submodule 922 is used to process the high-dimensional embedding vector in parallel, and captures the local fluctuation features of the sequence by concatenating the bidirectional hidden states of forward LSTM and backward LSTM.

[0162] The fusion prediction module 93 is used to concatenate the global trend features and local fluctuation features, and then dynamically weight and fuse them through a fully connected layer to generate a photovoltaic power generation prediction result for the future time period in one go; the training optimization module 94 is used to train the model with the mean square error as the loss function and the Adam optimizer, and retain the optimal generalization performance model by combining the early stopping mechanism.

[0163] The functions and methods of the above modules correspond to each other, and will not be repeated here.

[0164] In adaptability tests of photovoltaic power plants in different regions, this invention demonstrated excellent generalization ability:

[0165] A 50MW power plant in East China (frequent rain and fog): Test set R 2 =0.942, MAE=0.168, compared to the traditional LSTM model (R 2 =0.88) Accuracy improved by 7%;

[0166] A 20MW power plant in North China (under strong sunlight): Test set R 2 =0.963, MSE=0.041, prediction delay ≤0.5 seconds, meeting the real-time dispatch requirements of the power grid.

[0167] In summary, this invention solves the core pain points of existing photovoltaic power prediction by using a hybrid deep learning architecture and modular system design. It combines high accuracy, high efficiency, and high flexibility, and can be widely applied to power prediction scenarios of various photovoltaic power plants.

[0168] The following is for reference. Figure 10 It illustrates an electronic device suitable for implementing embodiments of the present invention (e.g., Figure 1 The diagram shows the structure of a computer device 600 (a server or terminal device). Figure 10 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0169] like Figure 10 As shown, the computer device 600 includes a central processing unit (CPU) 601 and a graphics processing unit (GPU) 602, which can perform various appropriate actions and processes according to programs stored in read-only memory (ROM) 603 or programs loaded from storage section 609 into random access memory (RAM) 604. The RAM 604 also stores various programs and data required for the operation of the device 600. The CPU 601, GPU 602, ROM 603, and RAM 604 are interconnected via a bus 605. An input / output (I / O) interface 606 is also connected to the bus 605.

[0170] The following components are connected to I / O interface 606: an input section 607 including a keyboard, mouse, etc.; an output section 608 including an LCD, speakers, etc.; a storage section 609 including a hard disk, etc.; and a communication section 610 including a network interface card, such as a LAN card or modem. The communication section 610 performs communication processing via a network such as the Internet. A drive 611 may also be connected to I / O interface 606 as needed. A removable medium 612, such as a hard disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 611 as needed so that computer programs read from it can be installed into storage section 609 as needed.

[0171] In particular, according to embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 610, and / or installed from removable medium 612. When the computer program is executed by central processing unit (CPU) 601 and graphics processing unit (GPU) 602, the functions defined in the methods of this invention are performed.

[0172] It should be noted that the computer-readable medium described in this invention can be a computer-readable signal medium, a computer-readable medium, or any combination thereof. A computer-readable medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor device, apparatus, or any combination thereof. More specific examples of a computer-readable medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution apparatus, device, or device. In this invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than a computer-readable medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution device, apparatus, or apparatus. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0173] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0174] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based devices that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0175] The modules described in the embodiments of the present invention can be implemented in software or hardware. The described modules can also be located in a processor.

[0176] In another aspect, the present invention also provides a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to perform the methods and steps described in the first aspect of the embodiments of the present invention.

[0177] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A photovoltaic power generation prediction method based on hybrid deep learning, characterized in that, The method includes the following steps: S1: Data preparation and multidimensional embedding: Collect historical photovoltaic power generation and related environmental data, and after standardization, fuse multidimensional time series features into a high-dimensional embedding vector through value embedding, location embedding and timestamp embedding. S2: Parallel feature extraction: including the following sub-steps S21 and S22; S21: The Informer module is used to process the high-dimensional embedding vector, and the ProbSparse self-attention mechanism is used to select the Top-k key queries to participate in the attention calculation, capture the long-range dependencies of the sequence, and output global trend features; S22: The high-dimensional embedding vector is processed in parallel using a BiLSTM module. The local fluctuation features of the sequence are captured by concatenating the bidirectional hidden states of the forward LSTM and backward LSTM. S3: Feature Fusion and Prediction Generation: After concatenating the global trend features and local fluctuation features, the photovoltaic power generation prediction results for future time periods are generated in one go through dynamic weighted fusion of the fully connected layer. S4: Model Training and Optimization: Using mean squared error as the loss function, the Adam optimizer is used to train the model, and an early stopping mechanism is combined to retain the model with the best generalization performance.

2. The photovoltaic power generation prediction method based on hybrid deep learning according to claim 1, characterized in that, The implementation of the ProbSparse self-attention mechanism in sub-step S21 includes: a. Perform a linear projection on the query vector Q and key vector K of the input sequence; b. By calculating the sparsity measure of the query vector. Filter the top-k key queries; where q i ,k j These are the i-th and j-th row vectors of the query matrix Q and the key matrix K, respectively; d k L is the dimension of the key vector. K The length of the key sequence; c. Calculate the sparse attention matrix based on the screening results, and gradually compress the sequence length by combining it with stratified distillation.

3. The photovoltaic power generation prediction method based on hybrid deep learning according to claim 1, characterized in that, The hidden layer dimension of the BiLSTM module in sub-step S22 is 256-512. The forward LSTM processes the sequence in ascending time order, and the backward LSTM processes the sequence in descending time order. The bidirectional hidden states are merged into a dimension 2·d through a concatenation operation. hidden Local fluctuation characteristics; where d hidden It is the hidden dimension of a single LSTM.

4. The photovoltaic power generation prediction method based on hybrid deep learning according to claim 1, characterized in that, The dynamic weighted fusion described in step S3 is implemented through a two-layer fully connected network. The first layer uses the GELU activation function, and the second layer is a linear mapping. The calculation formula is: y pred =W2·GEUL(W1·[F global ;F local ]+b1)+b2, where, [F global ;F local ] is the concatenated vector of global and local features, W1 and W2 are weight matrices, and b1 and b2 are bias terms.

5. The photovoltaic power generation prediction method based on hybrid deep learning according to claim 1, characterized in that, The one-time generation of photovoltaic power generation prediction results for future time periods mentioned in step S3 refers to directly outputting the complete prediction sequence for the next 0-24 hours through parallel computing, rather than generating it point by point through autoregression.

6. The photovoltaic power generation prediction method based on hybrid deep learning according to claim 1, characterized in that, The timestamp embedding in step S1 includes multi-granularity time features of year, month, day, hour, and minute, which are mapped to d through the FixedEmbedding layer. model Summation in 3D space.

7. The photovoltaic power generation prediction method based on hybrid deep learning according to claim 1, characterized in that, The model training described in step S4 adopts a teacher-forcing strategy, the decoder input is a combination of real sequence and zero mask, and the early stopping mechanism sets the patience to 10-20 epochs.

8. A photovoltaic power generation prediction system based on hybrid deep learning, characterized in that, include: The data processing and embedding module is used to collect historical photovoltaic power generation and related environmental data. After standardization, the multi-dimensional time series features are fused into a high-dimensional embedding vector through value embedding, location embedding and timestamp embedding. The parallel feature extraction module includes an Informer submodule and a BiLSTM submodule, wherein: The Informer submodule is used to process the high-dimensional embedding vector, selects Top-k key queries to participate in attention calculation through the ProbSparse self-attention mechanism, captures long-range dependencies of the sequence, and outputs global trend features; The BiLSTM submodule is used for parallel processing of the high-dimensional embedding vector. It captures the local fluctuation features of the sequence by concatenating the bidirectional hidden states of the forward LSTM and backward LSTM. The fusion prediction module is used to concatenate the global trend features and local fluctuation features, and then dynamically weight and fuse them through a fully connected layer to generate a photovoltaic power generation prediction result for a future time period in one go. The training optimization module is used to train the model with the mean squared error as the loss function and the Adam optimizer, and combines the early stopping mechanism to retain the model with the best generalization performance.

9. An electronic device, comprising: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the photovoltaic power generation prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the photovoltaic power generation prediction method as described in any one of claims 1 to 7.

Images