A distributed photovoltaic power generation prediction method and system based on lightweight feature migration

Abstract

The application discloses a kind of distributed photovoltaic power generation prediction method and system based on lightweight feature migration, method includes: source domain and target domain photovoltaic time series data are decomposed into high-frequency component and low-frequency component respectively, and segmented processing source domain and target domain high-low frequency component, wherein source domain low-frequency segment length is greater than the length of remaining segment;Source domain low-frequency segment is converted into target domain segment length by feature compression;Based on long short-term memory network, the low-frequency segment and high-frequency segment of source domain after compression are extracted for time series features respectively;Lightweight migration source domain high-low frequency final state feature initializes target domain long short-term memory network state vector respectively;Independent channel long short-term memory network is used to encode target domain high-frequency and low-frequency features respectively;Target domain high-low frequency features are spliced and photovoltaic power generation prediction results are output by using decoder.The application significantly improves the power prediction accuracy of newly-built distributed photovoltaic system, while having lower computational overhead.

Description

Technical Field

[0001] This invention relates to the field of distributed photovoltaic power generation forecasting, and specifically to a method and system for day-ahead forecasting of distributed photovoltaic power generation under small sample conditions. Background Technology

[0002] Distributed photovoltaic (PV) power generation forecasting exhibits significant temporal volatility due to its reliance on frequently fluctuating local micrometeorological conditions. Currently, there are three main methods for PV power generation forecasting: physical modeling, statistical methods, and machine learning methods. Physical modeling methods are based on the operating principles and physical characteristics of PV systems, combined with comprehensive meteorological data for prediction. However, this method requires support from ground-based sky imagery and weather station data, which are rarely installed near distributed PV systems. While satellite cloud imagery and numerical weather prediction can provide weather forecasts at different spatiotemporal scales, third-party data is costly and has relatively low accuracy for distributed PV forecasts.

[0003] Statistical learning methods, which typically rely solely on historical power generation data, are effective at modeling the autocorrelation of time series. However, these methods are limited in capturing complex nonlinear patterns. Machine learning methods excel at extracting complex sequential features from historical photovoltaic data, but newly installed photovoltaic systems often lack the large amounts of data required to support model training.

[0004] Transfer learning is a common method to address the limited data available for newly installed distributed photovoltaic (PV) systems. However, existing transfer learning methods suffer from high computational costs, require high inter-domain similarity, and involve complex feature extractor designs. Furthermore, when the sample size in the target domain is small, pre-training and fine-tuning methods may lead to negative transfer, affecting prediction performance. Summary of the Invention

[0005] Purpose of the invention: To address the shortcomings of existing technologies, this invention proposes a distributed photovoltaic power generation prediction method and system based on lightweight feature transfer, in order to solve the problems of high computational cost, high inter-domain similarity, complex feature extractor design, and the tendency to generate negative transfer effects on prediction performance when the sample size of the target domain is small.

[0006] Technical solution: The objective of this invention can be achieved through the following technical solution:

[0007] In a first aspect, the present invention provides a distributed photovoltaic power generation prediction method based on lightweight feature transfer, comprising the following steps:

[0008] (1) The photovoltaic time series data of the source domain and the target domain are decomposed into high-frequency components and low-frequency components respectively, and the high-frequency and low-frequency components of the source domain and the target domain are processed in segments, wherein the segment length of the low-frequency segment of the source domain is greater than the segment length of the other segments.

[0009] (2) The low-frequency segments of the source domain are compressed and converted into target domain segments. Temporal features are extracted from the low-frequency segments and high-frequency segments of the source domain after compression based on the long short-term memory network.

[0010] (3) Lightweight transfer of high and low frequency final state features from the source domain is used to initialize the target domain long short-term memory network state vector.

[0011] (4) Use an independent channel long short-term memory network to encode the high-frequency and low-frequency features of the target domain respectively;

[0012] (5) The high and low frequency features of the target domain are spliced ​​together and the photovoltaic power generation prediction results are output using the decoder.

[0013] In some preferred embodiments, in step (1), the photovoltaic time series data is decomposed into low-frequency components that characterize the main trend and periodicity of photovoltaic output and high-frequency components that characterize rapid fluctuations and transient characteristics using discrete wavelet transform.

[0014] In some preferred embodiments, in step (1), for source domain data, the length of the low-frequency segment is greater than that of the high-frequency segment, and for target domain data, the lengths of the high-frequency and low-frequency segments are the same.

[0015] In some preferred embodiments, in step (2), a linear fully connected network is used to independently reduce the dimensionality of each low-frequency segment of the source domain to match the segment length of the target domain.

[0016] In some preferred embodiments, in step (3), the hidden state vector and cell state vector of the source domain's long short-term memory network are used as the initial hidden state vector and cell state vector of the target domain's long short-term memory network.

[0017] In some preferred embodiments, in step (4), the final encoded states of the two channels are spliced ​​together and input into a first-order parallel long short-term memory network decoder with position embedding to capture the time dependency. Finally, the prediction results of all segments are sequentially connected to output the final photovoltaic power generation prediction result.

[0018] Secondly, the present invention provides a distributed photovoltaic power generation prediction system based on lightweight feature transfer, comprising:

[0019] The preprocessing module is used to decompose the photovoltaic time series data of the source domain and the target domain into high-frequency components and low-frequency components respectively, and process the high-frequency and low-frequency components of the source domain and the target domain in segments, wherein the segment length of the low-frequency segment of the source domain is greater than the segment length of the other segments;

[0020] The source domain feature compression and transfer module is used to compress the low-frequency segments of the source domain into the target domain segment length, extract temporal features from the compressed low-frequency segments and high-frequency segments of the source domain based on the long short-term memory network, and initialize the target domain long short-term memory network state vector with the final state features of the high and low frequencies of the source domain respectively.

[0021] The target domain prediction module is used to encode the high-frequency and low-frequency features of the target domain using an independent channel long short-term memory network; and to concatenate the high-frequency and low-frequency features of the target domain and output the photovoltaic power generation prediction result using a decoder.

[0022] Thirdly, the present invention provides a computer system including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the distributed photovoltaic power generation prediction method based on lightweight feature transfer.

[0023] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the distributed photovoltaic power generation prediction method based on lightweight feature transfer.

[0024] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned distributed photovoltaic power generation prediction method based on lightweight feature transfer.

[0025] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0026] (1) The photovoltaic data preprocessing strategy based on sequence decomposition and multi-scale segmentation of the present invention can accurately capture regularity and volatility characteristics. The design of the low-frequency segment length of the source domain being greater than the length of the other segments strengthens the extraction of long-term trend information in the low-frequency components and improves the accuracy of the prediction model in capturing long-term patterns.

[0027] (2) This invention establishes a source domain feature compression and lightweight transfer method, which extracts source domain features and initializes the target domain state vector based on long short-term memory network processing, thus solving the challenge of long sequence data transfer.

[0028] (3) The present invention is based on the target domain prediction model of independent channel long short-term memory network, and independently encodes high and low frequency components to avoid negative influence between heterogeneous features, thereby improving the prediction accuracy.

[0029] (4) Experiments show that the present invention can significantly improve the power generation prediction accuracy of newly built distributed photovoltaic systems, while having low computational overhead and is suitable for edge deployment. Attached Figure Description

[0030] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0031] Figure 2 This is a diagram showing the results of a day-ahead photovoltaic forecast under typical meteorological conditions in an embodiment of the present invention;

[0032] Figure 3 This is a graph showing the day-ahead prediction error results under different target domain photovoltaic scenarios in an embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] This invention discloses a distributed photovoltaic power generation prediction method based on lightweight feature transfer, mainly involving three aspects: a photovoltaic data preprocessing strategy based on multi-scale segmentation of sequence decomposition, source domain feature compression and lightweight transfer method, and a target domain prediction architecture based on independent channel long short-term memory network. Figure 1 As shown, the method mainly includes the following steps:

[0035] Step S1: Decompose the photovoltaic time series data of the source domain and the target domain into high-frequency components and low-frequency components respectively, and process the high and low frequency components of the source domain and the target domain in segments, wherein the segment length of the low frequency segment of the source domain is greater than the segment length of the other segments.

[0036] In some embodiments, step S1 can utilize discrete wavelet transform to decompose photovoltaic time-series data into low-frequency components characterizing the main trends and periodic patterns of photovoltaic output, and high-frequency components characterizing rapid fluctuations and transient characteristics. A multi-scale segmentation strategy is established for the source domain data, and long and short segmentation is performed according to the characteristics of different frequency components.

[0037] Step S2: The low-frequency segments of the source domain are compressed into target domain segments. Temporal features are extracted from the compressed low-frequency segments and high-frequency segments of the source domain based on the long short-term memory network.

[0038] Step S3: Lightweight transfer of high and low frequency final state features from the source domain to initialize the target domain long short-term memory network state vector.

[0039] Steps S2 and S3 achieve source domain feature compression and lightweight transfer. By compressing low-frequency features of the source domain, inter-domain differences are eliminated and redundant information is reduced. Based on the long short-term memory network, temporal analysis of the source domain is performed, important features are selectively transferred, and the generated feature state is lightweightly transferred to the target domain as the initial state of the target domain's long short-term memory network.

[0040] In some embodiments, by compressing the long segments of the low-frequency components of the photovoltaic time series in the source domain, a linear fully connected network can be used to independently reduce the dimensionality of each segment, remove redundant information, and match the segment length to the target domain segment length, thereby eliminating inter-domain differences and reducing migration computation overhead. A long short-term memory (LSTM) network is used to perform time series analysis on the source domain segments, and a gating mechanism is used to adjust the feature state vectors. The hidden state vectors and unit state vectors finally obtained from the LTM network in the source domain are used as the initial hidden state vectors and unit state vectors of the LTM network in the target domain, providing effective start-up information for the target domain model.

[0041] Step S4: Use an independent channel long short-term memory network to encode the high-frequency and low-frequency features of the target domain respectively.

[0042] Step S5: Concatenate the high and low frequency features of the target domain and use the decoder to output the photovoltaic power generation prediction result.

[0043] Steps S4 and S5 implement a target domain prediction architecture based on an independent channel long short-term memory network. Independent encoding of high and low frequency components can effectively avoid negative influences between heterogeneous features. In some embodiments, a single-stage parallel decoder can be used to generate multi-step prediction results to reduce error accumulation. For example, the high-frequency and low-frequency components of the target domain are encoded by two independent long short-term memory network channels respectively. The final encoded states of the two channels are concatenated and input into a single-stage parallel long short-term memory network decoder with positional embedding to capture temporal dependencies. Finally, the prediction results of all segments are sequentially concatenated to output the final photovoltaic power generation prediction result.

[0044] The following section provides a detailed explanation of a distributed photovoltaic power generation prediction method based on lightweight feature transfer, as disclosed in this invention, using specific sequence decomposition, segmented representation, and relevant formulas of the LSTM network model.

[0045] (1) Photovoltaic data preprocessing strategy based on multi-scale segmentation of sequence decomposition

[0046] For example, the original data is first decomposed using discrete wavelet transform, and then a multi-scale segmentation strategy is applied to different frequency components to specifically learn the overall trend and fluctuation characteristics of the photovoltaic output curve. This includes the following two aspects:

[0047] (1.1) Discrete Wavelet Decomposition. Directly using small samples of raw photovoltaic data is prone to model overfitting due to its large and frequent fluctuations. Photovoltaic data mainly consists of two components: low-frequency components caused by the regular changes in solar radiation due to the Earth's rotation, and high-frequency components caused by random cloud cover or rapid micro-meteorological fluctuations. To identify the regularity and volatility of photovoltaic output, discrete wavelet decomposition is used, defined as follows:

[0048]

[0049] Where y represents the discrete-time series of photovoltaic output, n is the discrete-time index, y[n] represents the power generation at time n, a and b are the scaling and translation parameters, respectively, the wavelet function ψ(·) is used to extract the frequency features in the photovoltaic time series, and q and k are the scaling and translation exponents, respectively. In this embodiment, the Daubechies-4 mother wavelet is used, and the number of decomposition layers is set to 3, which can effectively separate the high-frequency and low-frequency components in the photovoltaic sequence.

[0050] (1.2) Multi-scale segmentation strategy. Low-frequency components extract the main patterns of photovoltaic output, exhibiting smooth and stable characteristics; high-frequency components demonstrate fine-grained fluctuations, characterized by variability and irregularity. Based on the different characteristics of low-frequency and high-frequency components, and targeting the source domain photovoltaic time series Y... S A multi-scale segmentation strategy is proposed. Low-frequency components in the source domain. The sun's daily movement patterns are captured using relatively long segments, as follows:

[0051]

[0052] in, Each segment represents a low-frequency component, M is the number of low-frequency component data points in the source domain, and τ is the number of data points within a single long segment, set to an integer multiple of the daily photovoltaic data collection points. Through comprehensive analysis of multi-day data, the daily patterns of photovoltaic time series are captured.

[0053] Conversely, to fully capture the fluctuation characteristics of distributed photovoltaic time series, high-frequency components employ shorter segment lengths and higher time resolution. Source domain high-frequency components. The short-scale piecewise segmentation can be represented as:

[0054]

[0055] in, Each segment represents a high-frequency component, and δ (δ<τ) is the number of photovoltaic data points in a single short-scale segment to capture different fluctuation amplitudes and distributions within the photovoltaic power generation cycle.

[0056] For the low-frequency components of the target domain and high frequency components The following short-scale segmentation is adopted:

[0057]

[0058]

[0059] in, and Let N represent each segment of the low-frequency and high-frequency components of the target domain, respectively, where N is the number of data points in the target domain. Both low-frequency and high-frequency components are divided into N / δ segments to maximize the utilization of limited samples. Short-scale segmentation facilitates fine-grained analysis of recent features in the target domain and, compared to no segmentation, reduces the gradient flow path length during training, avoiding gradient vanishing or exploding. Furthermore, segmentation of the target domain requires more iterations, implicitly increasing the weights during the learning process. This strategy balances data differences between domains and mitigates catastrophic forgetting in Long Short-Term Memory networks.

[0060] (2) Source domain feature compression and lightweight transfer method

[0061] Lightweight feature transfer methods aim to streamline the source domain transfer process while preserving key information. First, feature compression is applied to long segments of the source domain to eliminate redundant features and reduce inter-domain differences. Then, a long short-term memory network gating mechanism is used to obtain comprehensive features of the source domain, achieving lightweight feature transfer. Specifically, this includes the following two aspects:

[0062] (2.1) Source Domain Feature Compression. The source domain extracts generalized features of the overall intraday trend of photovoltaic power generation using long-scale segmented low-frequency sequences. However, compared to high-frequency irregular components, low-frequency components exhibit more obvious patterns and are easier to learn. Directly inputting long-segment sequences into the Long Short-Term Memory (LSTM) network increases complexity and distracts from valuable features, negatively impacting performance. Therefore, long-segment sequences from the source domain... Use the following feature compression:

[0063]

[0064] Each source domain long sequence is processed by the function F. C (·) Features are obtained by compression alone. F C (·) Implemented through a single-layer linear fully connected network, it achieves good feature extraction performance without introducing high complexity. After feature compression, the long-scale segment length of the source domain is converted into the segment length δ of the target domain, standardizing the feature dimension and promoting seamless integration of cross-domain information, enabling the model to focus on learning domain-invariant features. Feature compression achieves a balance between efficiency and effectiveness, providing meaningful feature representations for subsequent modeling.

[0065] (2.2) Lightweight Feature Transfer Method Based on Long Short-Term Memory Networks. This method utilizes a gated Long Short-Term Memory network architecture to control the information flow between the source and target domains. Lightweight feature transfer focuses on propagating important states. The update mechanism for the hidden states and unit states of the Long Short-Term Memory network is as follows:

[0066]

[0067]

[0068]

[0069] Among them, W C and b C These are learnable parameters. Candidate cell state, cell state As a long-term memory component, the hidden state accumulates valuable features from the source domain through forgetting and input gates. As a feature carrier, the output gate encodes in The knowledge of photovoltaic sources is transferred to the current time step. and By transmitting features containing source domain synthesis characteristics and Enables lightweight migration from the source domain to the target domain.

[0070] (3) Target domain prediction architecture based on independent channel long short-term memory network

[0071] A prediction model based on independent-channel Long Short-Term Memory (LSTM) networks enables efficient processing of target domain data. First, time-series analysis of heterogeneous photovoltaic data is performed using independent-channel LSTM network encoders. Then, the outputs of the two channels are concatenated and input into a single-pass, multi-step LSTM network decoder with positional embeddings to capture temporal dependencies and accelerate computation through parallel processing. Specifically, this includes the following two aspects:

[0072] (3.1) Independent Channel Long Short-Term Memory Network Encoder. Low-frequency and high-frequency components of photovoltaic data have different characteristics and generation mechanisms. Features with weak correlation are prone to mutual interference and negative impacts when processed together. Furthermore, for target domains with small datasets, feature interactions increase model complexity and exacerbate the risk of overfitting. Therefore, an independent channel long short-term memory network encoder for the target domain is proposed. This encoder processes the pre-processed target domain data... and The input consists of two independent channels, whose initial states are the same as the final states of the source domain, respectively. and The two parallel channels of the target domain independent channel long short-term memory network encoder are represented as follows:

[0073]

[0074]

[0075] Among them, Channel L (·) and Channel H (·) represent the low-frequency and high-frequency channels using the encoder, respectively. Each channel independently encodes different photovoltaic frequency components, and parallel processing accelerates computation. The final hidden state from both channels. and unit state By splicing together the two frequency characteristics, a representation of short-term fluctuations and long-term patterns in photovoltaic data is provided.

[0076] (3.2) Single-Step Multi-Step Long Short-Term Memory Network Decoder. Traditional multi-step photovoltaic prediction long short-term memory network decoders employ a recursive prediction method, where the prediction at each step is used as input for subsequent predictions until the entire photovoltaic sequence is generated. This method results in error accumulation and slow prediction speed. To overcome the limitations of traditional multi-step photovoltaic prediction long short-term memory network decoders, this embodiment also designs a single-step multi-step long short-term memory network decoder. This decoder divides the prediction time domain into p = n... p / δ time periods, where n p Let be the number of time steps to be predicted, and δ represent the width of each prediction segment. Furthermore, considering the differences in temporal dependencies exhibited by different segments in the photovoltaic sequence—for example, adjacent segments typically show strong correlations while distant segments have weaker relationships—a learnable location embedding pe∈R is introduced. p×d Where d represents the embedding dimension, the target domain photovoltaic data and the location embedding are concatenated to form the complete input of the decoder, enabling the decoder to independently identify photovoltaic-specific temporal features. When photovoltaic data is limited, the location embedding concatenation design better preserves the two information streams compared to additive embedding. The decoder's initial state is the encoder's final concatenation state. and It generates predictions for multiple future time steps simultaneously based on a linear layer, and finally arranges them in chronological order to generate a prediction sequence.

[0077] In summary, this invention discloses a distributed photovoltaic (PV) power generation prediction method based on lightweight feature transfer. This method preprocesses the PV time series data through discrete wavelet decomposition and multi-scale segmentation, effectively separating low-frequency and high-frequency features. A lightweight feature transfer algorithm is employed to achieve feature compression and adaptive training, reducing computational complexity. An independent dual-channel long short-term memory (LSTM) network model is designed to process different frequency components separately, and a one-time decoder is used to generate prediction results, effectively preventing negative interference between features and reducing error accumulation. Compared to existing transfer learning methods, this embodiment's method has lower computational cost, better transfer performance, and stronger adaptability, making it particularly suitable for power generation prediction of newly installed distributed PV systems.

[0078] To further illustrate the effectiveness of this invention, this embodiment selects a distributed photovoltaic array in Alice Springs, Australia, for case analysis. Photovoltaic time series data were collected between 05:00 and 20:00 local time, with a collection interval of 15 minutes. Each experiment used 30 days of target domain data, divided into training and test sets at an 8:2 ratio. The source domain used data from January 1 to June 30, 2023. The target domain employed a grouped time series cross-validation method to conduct experiments on data from four seasons, with each season divided into 5 groups. A sliding window was used to sequentially select individual experimental data groups, with an interval of 6 days between adjacent groups (test set length). The specific time periods selected for the target domain are as follows: Winter (July 1 to August 23, 2023), Spring (October 1 to November 23, 2023), Summer (January 1 to February 23, 2024), and Autumn (April 1 to May 24, 2024).

[0079] In addition, to verify the transferability of the proposed method, six distributed photovoltaic arrays were selected and labeled as PV1 to PV6, with PV1 serving as the source domain and the other photovoltaic arrays serving as the target domain. These photovoltaic arrays have significant differences in rated power, array structure, etc. The hardware parameters of each photovoltaic array are detailed in Table 1.

[0080] Table 1 Specific hardware parameters for each photovoltaic system

[0081] Rated power Photovoltaic model Array structure Material PV 1 10.5kW Trina TSM-175DC01 Dual-axis tracking Monocrystalline silicon PV 2 6.96kW First Solar FS-272* Ground fixed type Cadmium telluride PV 3 6kW Kaneka G-EA060 Ground fixed type amorphous silicon PV 4 4.95kW BP 3165 Roof-mounted Polycrystalline silicon PV 5 5.4kW Kyocera KD135GX-LP Single-axis tracking Polycrystalline silicon PV 6 5.25kW Trina TSM-175DC01 Ground fixed type Monocrystalline silicon

[0082] This experiment uses two widely used evaluation metrics: mean absolute error (MAE) and root mean square error (RMSE), and their calculation formulas are shown below:

[0083]

[0084]

[0085] Where, N test y represents the number of test datasets. predFor the predicted value, y true This represents the actual value.

[0086] The prediction error assessment results of photovoltaic PV6 under different seasons obtained by the method disclosed in this application are shown in Table 2.

[0087] Table 2 Prediction error of PV6 in different seasons

[0088] MAE RMSE spring 0.4127 0.6164 summer 0.3955 0.5878 autumn 0.1763 0.3347 winter 0.2196 0.4111

[0089] The evaluation results show that the method proposed in this application can ensure a reasonable prediction error range under different seasonal conditions, adapt to the weather-dependent changes in photovoltaic power generation, and ensure that reliable prediction results can still be provided under variable weather conditions.

[0090] like Figure 2 The specific prediction results show that the method of this application exhibits high prediction accuracy under different typical meteorological conditions (sunny, cloudy, and rainy). Under sunny conditions, the predicted values ​​are basically consistent with the actual values; under cloudy and rainy conditions, despite the complex and variable meteorological conditions, the prediction results can still maintain a reasonable error range, demonstrating good adaptability.

[0091] The prediction error assessment results of photovoltaics in different target domains in spring using the method disclosed in this application are shown in Table 3.

[0092] Table 3 Prediction errors of photovoltaics in different target domains

[0093] MAE RMSE PV 2 0.3303 0.5410 PV 3 0.4614 0.7468 PV 4 0.5198 0.8536 PV 5 0.4754 0.7642 PV 6 0.4127 0.6164

[0094] The evaluation results show that the method of this application exhibits low prediction error in all target domains, remaining within a reasonable range, indicating that the method demonstrates good generalization ability in different target domains.

[0095] like Figure 3 As shown, even with a small amount of training data in the target domain, the proposed method still maintains a low prediction error. With an increase in the amount of training data, the prediction error further decreases and tends to stabilize, demonstrating that the method still possesses high prediction accuracy and robustness even with limited sample size. This result indicates that the proposed method can still provide reliable prediction results under conditions of scarce data, fully demonstrating its superiority and adaptability in small-sample photovoltaic time-series prediction.

[0096] Based on the same inventive concept, this invention also discloses a distributed photovoltaic power generation prediction system based on lightweight feature transfer, comprising: a preprocessing module, used to decompose the source domain and target domain photovoltaic time-series data into high-frequency components and low-frequency components respectively, and process the source domain and target domain high and low frequency components in segments, wherein the low-frequency segment length of the source domain is greater than the length of the other segments; a source domain feature compression and transfer module, used to compress the source domain low-frequency segments into target domain segment lengths, and extract time-series features from the compressed low-frequency segments and high-frequency segments of the source domain based on a long short-term memory network; and, the lightweight transfer of the source domain high and low frequency final state features to initialize the target domain long short-term memory network state vectors respectively; a target domain prediction module, used to encode the target domain high-frequency and low-frequency features respectively using an independent channel long short-term memory network; and, the target domain high and low frequency features are concatenated and the photovoltaic power generation prediction result is output using a decoder. For details of the specific implementation of each module, please refer to the foregoing method implementations, which will not be repeated here.

[0097] This invention also discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the distributed photovoltaic power generation prediction method based on lightweight feature transfer.

[0098] This invention also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the distributed photovoltaic power generation prediction method based on lightweight feature transfer.

[0099] This invention also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the distributed photovoltaic power generation prediction method based on lightweight feature transfer.

[0100] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A distributed photovoltaic power generation prediction method based on lightweight feature transfer, characterized in that, Includes the following steps: (1) Decompose the photovoltaic time series data of the source domain and the target domain into high-frequency components and low-frequency components respectively, and process the high and low frequency components of the source domain and the target domain in segments, wherein the segment length of the low frequency segment of the source domain is greater than the segment length of the other segments; (2) The low-frequency segments of the source domain are compressed and converted into target domain segments. Temporal features are extracted from the low-frequency segments and high-frequency segments of the source domain after compression based on the long short-term memory network. (3) Lightweight transfer of the high and low frequency final state features of the source domain to initialize the state vector of the target domain long short-term memory network respectively; wherein the hidden state vector and the unit state vector of the source domain long short-term memory network are used as the initial hidden state vector and the unit state vector of the target domain long short-term memory network. (4) Use independent channel long short-term memory networks to encode the high-frequency and low-frequency features of the target domain respectively; (5) The high and low frequency features of the target domain are spliced ​​together and the photovoltaic power generation prediction results are output using the decoder.

2. The distributed photovoltaic power generation prediction method based on lightweight feature transfer according to claim 1, characterized in that, In step (1), the photovoltaic time series data is decomposed into low-frequency components that characterize the main trend and periodicity of photovoltaic output and high-frequency components that characterize rapid fluctuations and transient characteristics using discrete wavelet transform.

3. The distributed photovoltaic power generation prediction method based on lightweight feature transfer according to claim 1, characterized in that, In step (1), for source domain data, the length of the low-frequency segment is greater than that of the high-frequency segment, and for target domain data, the lengths of the high-frequency and low-frequency segments are the same.

4. The distributed photovoltaic power generation prediction method based on lightweight feature transfer according to claim 1, characterized in that, In step (2), a linear fully connected network is used to independently reduce the dimensionality of each low-frequency segment of the source domain to match the segment length of the target domain.

5. The distributed photovoltaic power generation prediction method based on lightweight feature transfer according to claim 1, characterized in that, In step (4), the final encoded states of the two channels are concatenated and input into a parallel long short-term memory network decoder with position embedding to capture the time dependency. Finally, the prediction results of all segments are sequentially connected to output the final photovoltaic power generation prediction result.

6. A distributed photovoltaic power generation prediction system based on lightweight feature transfer, characterized in that, include: The preprocessing module is used to decompose the photovoltaic time series data of the source domain and the target domain into high-frequency components and low-frequency components respectively, and process the high-frequency and low-frequency components of the source domain and the target domain in segments, wherein the segment length of the low-frequency segment of the source domain is greater than the segment length of the other segments; The source domain feature compression and transfer module is used to compress the low-frequency segments of the source domain into target domain segment lengths, and extract temporal features from the compressed low-frequency segments and high-frequency segments of the source domain based on a long short-term memory network. Furthermore, the high and low frequency final state features of the source domain are used to initialize the state vector of the target domain long short-term memory network, respectively; wherein the hidden state vector and unit state vector of the source domain long short-term memory network are used as the initial hidden state vector and unit state vector of the target domain long short-term memory network. The target domain prediction module is used to encode the high-frequency and low-frequency features of the target domain using an independent channel long short-term memory network; and to concatenate the high-frequency and low-frequency features of the target domain and output the photovoltaic power generation prediction result using a decoder.

7. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of a distributed photovoltaic power generation prediction method based on lightweight feature transfer as described in any one of claims 1-5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of a distributed photovoltaic power generation prediction method based on lightweight feature transfer as described in any one of claims 1-5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of a distributed photovoltaic power generation prediction method based on lightweight feature transfer as described in any one of claims 1-5.

Images