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

Abstract

The invention discloses a distributed photovoltaic power generation prediction method and system based on lightweight feature migration, and the method comprises the steps: decomposing photovoltaic time sequence data of a source domain and a target domain into a high-frequency component and a low-frequency component, and carrying out the segmentation processing of the high-frequency component and the low-frequency component of the source domain and the target domain, and enabling the length of the low-frequency segment of the source domain to be larger than the length of the other segments; performing feature compression on the low-frequency segments of the source domain to convert the low-frequency segments of the source domain into target domain segments, and performing time sequence feature extraction on the low-frequency segments and the high-frequency segments after the source domain compression based on a long-short-term memory network; respectively initializing target domain long and short-term memory network state vectors according to lightweight migration source domain high and low frequency final state characteristics; respectively coding high-frequency and low-frequency characteristics of a target domain by adopting an independent channel long-short-term memory network; and splicing the high and low frequency features of the target domain and outputting a photovoltaic power generation prediction result by using a decoder. According to the method, the power generation power prediction precision of the newly-built distributed photovoltaic system is remarkably improved, and meanwhile, the calculation overhead is relatively low.

Description

Technical Field

[0001] The present invention relates to the field of distributed photovoltaic power generation prediction, and in particular to a method and system for performing day-ahead prediction on distributed photovoltaic power generation under a small sample condition. Background Art

[0002] Distributed photovoltaic power generation forecasting shows obvious temporal volatility because it relies on frequently fluctuating local micro-meteorological conditions. At present, there are three main methods for photovoltaic power generation forecasting: physical model method, statistical method and machine learning method. The physical model method is based on the operating principle and physical characteristics of the photovoltaic system, combined with comprehensive meteorological data for forecasting. However, this method requires support from ground sky images and meteorological station data, and these devices are rarely installed near distributed photovoltaic systems. Although satellite cloud images and numerical weather forecasts can provide weather forecasts at different time and space scales, the cost of third-party data is high and the prediction accuracy of distributed photovoltaics is low.

[0003] Statistical learning methods usually rely only on historical power generation data and can effectively model the autocorrelation of time series. However, these methods are limited in capturing complex nonlinear patterns. Machine learning methods perform well in extracting complex sequence features from historical PV data, but newly installed PV systems often lack the large amount of data needed to support model training.

[0004] Transfer learning is a common method to solve the problem of limited data for newly installed distributed photovoltaic systems. However, existing transfer learning methods have problems such as high computational cost, high inter-domain similarity, and complex feature extractor design. In addition, when the sample size of the target domain is small, pre-training and fine-tuning methods may lead to negative transfer and affect the prediction effect. Summary of the invention

[0005] Purpose of the invention: In view of the shortcomings of the prior art, the present invention proposes a distributed photovoltaic power generation prediction method and system based on lightweight feature migration to solve the problems of the existing transfer learning methods in the above-mentioned background technology, such as high computational cost, high inter-domain similarity requirement, complex feature extractor design, and easy negative migration affecting the prediction effect when the target domain sample size is small.

[0006] Technical solution: The purpose of the present 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 migration, comprising the following steps:

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

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

[0010] (3) Lightweight migration of high- and low-frequency final state features of the source domain to respectively initialize the state vector of the long short-term memory network in the target domain;

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

[0012] (5) The high-frequency and low-frequency features of the target domain are concatenated and the decoder is used to output the photovoltaic power generation prediction results.

[0013] In some preferred embodiments, in step (1), discrete wavelet transform is used to decompose photovoltaic time series data into low-frequency components that characterize the main trends and periodic laws of photovoltaic output and high-frequency components that characterize rapid fluctuations and transient characteristics.

[0014] In some preferred embodiments, in step (1), for source domain data, the length of low frequency segments is greater than that of high frequency segments, and for target domain data, the lengths of high 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 dimension of each low-frequency segment in 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 unit state vector finally obtained from the long short-term memory network of the source domain are used as the initial hidden state vector and unit state vector of the long short-term memory network of the target domain.

[0017] In some preferred embodiments, 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 temporal dependency, and finally all segmented prediction results are sequentially connected to output the final photovoltaic power generation prediction result.

[0018] In a second aspect, the present invention provides a distributed photovoltaic power generation prediction system based on lightweight feature migration, comprising:

[0019] A preprocessing module is 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 length of the source domain low-frequency segment is greater than the length of the other segments;

[0020] The source domain feature compression and migration module is used to convert the source domain low-frequency segmentation into the target domain segment length through feature compression, and extract the temporal features of the source domain compressed low-frequency segmentation and high-frequency segmentation based on the long short-term memory network; and, the lightweight migration of the source domain high- and low-frequency final state features respectively initializes the target domain long short-term memory network state vector;

[0021] The target domain prediction module is used to encode the high-frequency and low-frequency features of the target domain respectively using independent channel long short-term memory networks; 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] In a third aspect, the present invention provides a computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of a distributed photovoltaic power generation prediction method based on lightweight feature migration are implemented.

[0023] In a fourth aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the distributed photovoltaic power generation prediction method based on lightweight feature migration are implemented.

[0024] In a fifth aspect, the present invention provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of a distributed photovoltaic power generation prediction method based on lightweight feature migration.

[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 in the present invention can accurately capture the regularity and volatility characteristics. The design of the source domain low-frequency segment length being longer than the other segment lengths strengthens the extraction of long-term trend information of the low-frequency component and improves the accuracy of the prediction model in capturing long-term laws.

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

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

[0029] (4) Experiments show that the present invention can significantly improve the prediction accuracy of power generation of newly built distributed photovoltaic systems, while having low computational overhead and being suitable for edge deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a method flow chart of an embodiment of the present invention;

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

[0032] Figure 3 1 is a diagram of the day-ahead prediction error results under different target domain photovoltaics in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

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

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

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

[0037] Step S2: feature compress the low-frequency segments in the source domain and convert them into the target domain segment length, and extract temporal features from the low-frequency segments and high-frequency segments compressed in the source domain based on the long short-term memory network.

[0038] Step S3: lightweight migration of 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.

[0039] Steps S2 and S3 implement source domain feature compression and lightweight migration. They eliminate inter-domain differences and reduce redundant information by compressing low-frequency features in the source domain. They perform temporal analysis on the source domain based on the long short-term memory network, selectively transfer important features, and lightweight migrate the generated feature state to the target domain as the initial state of the long short-term memory network in the target domain.

[0040] In some embodiments, by compressing the long segments of the low-frequency components of the source domain photovoltaic time series, a linear fully connected network can be used to independently reduce the dimension of each segment, remove redundant information and match the length of the target domain segment, eliminate the difference between domains and reduce the migration calculation overhead. The source domain segment is analyzed in time series by a long short-term memory network, and the characteristic state vector is adjusted by a gating mechanism. The hidden state vector and unit state vector finally obtained by the long short-term memory network of the source domain are used as the initial hidden state vector and unit state vector of the long short-term memory network of the target domain, providing effective startup information for the target domain model.

[0041] In step S4, independent channel long short-term memory networks are used 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 the target domain prediction architecture based on independent channel long short-term memory networks. Independently encoding high and low frequency components can effectively avoid negative effects between heterogeneous features. In some embodiments, a single parallel decoder can be used to generate multi-step prediction results to reduce error accumulation. For example, two independent long short-term memory network channels are used to encode the high-frequency and low-frequency components of the target domain, respectively, and the final encoding states of the two channels are spliced ​​and input into a single parallel long short-term memory network decoder with position embedding to capture time dependencies. Finally, all segmented prediction results are sequentially connected to output the final photovoltaic power generation prediction results.

[0044] In the following, a distributed photovoltaic power generation prediction method based on lightweight feature migration disclosed in an embodiment of the present invention is described in detail in combination with specific sequence decomposition, segmented representation and LSTM network model related formulas.

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

[0046] For example, the original data is first decomposed by 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. Specifically, it includes the following two aspects:

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

[0048]

[0049] Among them, 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 scale parameters and translation parameters respectively, the wavelet function ψ(·) is used to extract the frequency characteristics in the photovoltaic time series, and q and k are scale index and translation index respectively. This embodiment adopts the Daubechies-4 mother wavelet, and the decomposition level 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. The low-frequency component extracts the main law of photovoltaic output and has the characteristics of smoothness and stability; the high-frequency component reflects fine-grained fluctuations and has the characteristics of changeability and irregularity. According to the different characteristics of low-frequency and high-frequency components, the source domain photovoltaic time series Y S A multi-scale segmentation strategy is proposed. Low-frequency components in the source domain A longer segment length is used to capture the daily motion of the sun, divided into the following segments:

[0051]

[0052] in, Represents each segment of the low-frequency component, M is the number of data points of the low-frequency component in the source domain, and τ is the number of data points in a single long segment and is set to an integer multiple of the number of data points collected by daily photovoltaics. Through the comprehensive analysis of multi-day data, the daily regularity of the photovoltaic time series is captured.

[0053] On the contrary, in order to fully capture the fluctuation characteristics of distributed photovoltaic time series, the high-frequency component adopts a shorter segment length and a higher time resolution. The short-scale segmentation of can be expressed as:

[0054]

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

[0056] For the low-frequency components of the target domain and high frequency components The following short-scale segments are used:

[0057]

[0058]

[0059] in, and Represents each segment of the low-frequency component and high-frequency component of the target domain, and N is the number of target domain data points. Both low-frequency and high-frequency components are divided into N / δ segments to maximize the utilization of limited samples. Short-scale segmentation is conducive to fine-grained analysis of recent features of the target domain, and compared with no segmentation, it reduces the length of the gradient flow path during training and avoids gradient vanishing or explosion. In addition, the segmentation of the target domain takes up more iterations, implicitly increasing the weight in the learning process. This strategy balances the data differences between domains and alleviates catastrophic forgetting in long short-term memory networks.

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

[0061] The lightweight feature transfer method is committed to lightweight source domain transfer process while retaining key information. First, feature compression is applied to the long segments of the source domain to eliminate redundant features and reduce the difference between domains. Then, the comprehensive features of the source domain are obtained based on the long short-term memory network gating mechanism to achieve lightweight feature transfer. Specifically, it includes the following two aspects:

[0062] (2.1) Source domain feature compression. The source domain extracts generalized features of the overall trend of photovoltaic power generation within a day through long-scale segmented low-frequency sequences. However, compared with high-frequency irregular components, low-frequency components have more obvious patterns and are easier to learn. Directly inputting long-segment sequences into the long short-term memory network will increase complexity and distract attention from valuable features, which will have a negative impact on performance. Therefore, long-segment sequences in the source domain are compressed. Use feature compression as shown below:

[0063]

[0064] Among them, each source domain long segment sequence is passed through the function F C (·) Features obtained by compression alone F C (·) It is implemented through a single-layer linear fully connected network, which has good feature extraction effect without introducing high complexity. After feature compression, the source domain long-scale segment length is converted to the target domain segment length δ, which standardizes the feature dimension and promotes the seamless connection of cross-domain information, allowing the model to focus on learning domain-invariant features. Feature compression strikes a balance between efficiency and effect, providing meaningful feature representation for subsequent modeling.

[0065] (2.2) Lightweight feature transfer method based on long short-term memory network. Based on the long short-term memory network architecture with a gating mechanism to control the information flow between the source domain and the target domain, lightweight feature transfer focuses on propagating important states. The update mechanism of the hidden state and unit state of the long short-term memory network is as follows:

[0066]

[0067]

[0068]

[0069] Among them, W C and b C is a learnable parameter, is the candidate unit state, the unit state As a long-term memory component, it accumulates valuable features in the source domain through the forget gate and the input gate. As a feature carrier, the output gate will be encoded in The source photovoltaic knowledge in is transferred to the current time step, and By transferring the comprehensive features of the source domain and Realize lightweight migration from source domain to target domain.

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

[0071] The long short-term memory network prediction model based on independent channels realizes the effective processing of target domain data. First, the long short-term memory network encoder of independent channels is used to perform time series analysis on heterogeneous photovoltaics respectively; then, the outputs of the two channels are spliced ​​and input into the single multi-step long short-term memory network decoder with position embedding to capture the time series dependency and parallel processing to speed up the calculation. It specifically includes the following two aspects:

[0072] (3.1) Independent channel long short-term memory network encoder. The low-frequency and high-frequency components of photovoltaic data have different characteristics and generation mechanisms. When the features with weak correlation are processed together, they are easy to interfere with each other and produce negative effects. Moreover, for the target domain with a small data set, feature interaction will increase the complexity of the model and aggravate the risk of overfitting. Therefore, a target domain independent channel long short-term memory network encoder is proposed. This encoder converts the preprocessed target domain data into and Input two independent channels, whose initial states are the final states of the source domain and The two parallel channels of the target domain independent channel long short-term memory network encoder are expressed as follows:

[0073]

[0074]

[0075] Among them, Channel L (·) and Channel H (·) denotes the low-frequency and high-frequency channels of the encoder, respectively. Each channel encodes the features of different photovoltaic frequency components independently, and the calculation speed is accelerated by parallel processing. The final hidden state from the two channels and cell status The two frequency characteristics are combined by splicing to provide short-term fluctuations and long-term regularities of photovoltaic data.

[0076] (3.2) Single multi-step LSTM decoder. The traditional multi-step photovoltaic prediction LSTM decoder adopts a recursive prediction method, that is, the prediction of each step is used as the input of the subsequent prediction until the entire photovoltaic sequence is generated. This method causes error accumulation and slow prediction speed. In order to overcome the limitations of the traditional multi-step photovoltaic prediction LSTM decoder, this embodiment also designs a single multi-step LSTM decoder. The decoder divides the prediction time domain into p = n p / δ periods, where n p is the number of time steps to be predicted, and δ represents the width of each predicted segment. In addition, in view of the differences in temporal dependencies between different segments in the PV sequence, for example, adjacent segments usually show strong correlations, while the relationship between distant segments is weak, a learnable position embedding pe∈R is introduced p×d , where d represents the embedding dimension. The target domain PV data and position embedding are concatenated to form the complete input of the decoder, enabling the decoder to independently identify PV-specific temporal features. In the case of limited PV data, the position embedding concatenation design can better preserve the two information streams than additive embedding. The initial state of the decoder is the final concatenated state of the encoder. and And based on the linear layer, it generates predictions for multiple future time steps at the same time, and finally arranges them in chronological order to generate a prediction sequence.

[0077] In summary, an embodiment of the present invention discloses a distributed photovoltaic power generation prediction method based on lightweight feature migration. This method pre-processes photovoltaic time series through discrete wavelet decomposition and multi-scale segmentation to effectively separate low-frequency and high-frequency features; a lightweight feature migration algorithm is used to achieve feature compression and adaptive training to reduce computational complexity; an independent dual-channel long short-term memory network model is designed to process different frequency components respectively, and a one-time decoder is used to generate prediction results, effectively preventing negative interference between features and reducing error accumulation. Compared with existing transfer learning methods, the method of this embodiment has smaller computational complexity, better migration effect, and stronger adaptability, and is particularly suitable for power generation prediction of newly installed distributed photovoltaic systems.

[0078] In order to further illustrate the effect of the present invention, this embodiment selects the distributed photovoltaic array in Alice Springs, Australia for example analysis, collects photovoltaic time series between 05:00 and 20:00 local time, and the collection interval is 15 minutes. Each experiment uses 30 days of target domain data, and is divided into training set and test set at a ratio of 8:2. The source domain uses fixed data from January 1, 2023 to June 30 of the same year. The target domain uses the grouped time series cross-validation method to experiment with four seasonal data respectively. Each season is divided into 5 groups of experiments. The sliding window sequence is used to select a single group of experimental data. The interval between two adjacent groups is 6 days (test set length). The specific time periods selected by 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, marked as PV1 to PV 6, where PV 1 was used as the source domain and the remaining photovoltaic arrays were used as target domains. 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 of each photovoltaic

[0081] Rated Power PV 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 Cadmium Telluride PV 3 6kW Kaneka G-EA060 Ground Fixed Amorphous Silicon PV 4 4.95kW BP 3165 Roof Fixed Polycrystalline Silicon PV 5 5.4kW Kyocera KD135GX-LP Single-axis Tracking Polycrystalline Silicon PV 6 5.25kW Trina TSM-175DC01 Ground Fixed Monocrystalline Silicon

[0082] This experiment adopts two widely used evaluation indicators, namely, mean absolute error (MAE) and root mean square error (RMSE), which are calculated as follows:

[0083]

[0084]

[0085] Among them, N test Indicates the number of test data sets, y predis the predicted value, y true is the actual value.

[0086] The prediction error evaluation results of photovoltaic PV6 in 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 proposed method can ensure a reasonable prediction error range under different seasonal conditions, adapt to the changes in the meteorological dependence of photovoltaic power generation, and ensure that reliable prediction results can be provided under variable meteorological conditions.

[0090] like Figure 2 As shown in the specific prediction results, the method of this application shows high prediction accuracy under different typical meteorological conditions (sunny, cloudy, rainy). Under sunny conditions, the predicted value is basically consistent with the actual value; under cloudy and rainy conditions, although the meteorological conditions are complex and changeable, the prediction results can still maintain a reasonable error range, showing good adaptability.

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

[0092] Table 3 Prediction error of photovoltaic 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 the present application shows low prediction errors in each target domain and remains within a reasonable range, indicating that the method shows good generalization in different target domains.

[0095] like Figure 3 As shown in the figure, when the amount of training data in the target domain is small, the method of this application can still maintain a low prediction error. As the amount of training data increases, the prediction error further decreases and tends to be stable, showing that the method still has high prediction accuracy and robustness when the sample size is limited. This result shows that the method of this application can still provide reliable prediction results under the condition of scarce data, which fully demonstrates its superiority and adaptability in small sample photovoltaic time series prediction.

[0096] Based on the same inventive concept, the embodiment of the present invention also discloses a distributed photovoltaic power generation prediction system based on lightweight feature migration, including: a preprocessing module, which is 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 source domain low-frequency segment length is greater than the other segment lengths; a source domain feature compression and migration module, which is used to convert the source domain low-frequency segment into the target domain segment length by feature compression, and extract the time series features of the source domain compressed low-frequency segment and high-frequency segment respectively based on the long short-term memory network; and, the lightweight migration source domain high- and low-frequency final state features respectively initialize the target domain long short-term memory network state vector; a target domain prediction module, which is 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 spliced ​​and the decoder is used to output the photovoltaic power generation prediction result. The specific module embodiments are described in detail in the aforementioned method embodiments, which will not be repeated.

[0097] An embodiment of the present 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, the steps of a distributed photovoltaic power generation prediction method based on lightweight feature migration are implemented.

[0098] An embodiment of the present invention further discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the distributed photovoltaic power generation prediction method based on lightweight feature migration are implemented.

[0099] The embodiment of the present invention further discloses a computer program product, including a computer program, which implements the steps of the distributed photovoltaic power generation prediction method based on lightweight feature migration when executed by a processor.

[0100] The above shows and describes 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, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. A distributed photovoltaic power generation prediction method based on lightweight feature migration, characterized in that: The following steps are involved: (1) The source domain and target domain photovoltaic time series data are decomposed into high-frequency components and low-frequency components, respectively, and the source domain and target domain high-frequency and low-frequency components are processed in segments, where the length of the source domain low-frequency segment is greater than the length of the other segments; (2) The low-frequency segments in the source domain are compressed and converted into target domain segments. The temporal features of the compressed low-frequency segments and high-frequency segments in the source domain are extracted based on the long short-term memory network. (3) Lightweight migration of high- and low-frequency final state features of the source domain to respectively initialize the state vector of the long short-term memory network in the target domain; (4) Using independent channel long short-term memory networks to encode high-frequency and low-frequency features of the target domain respectively; (5) The high-frequency and low-frequency features of the target domain are concatenated and the decoder is used to output the photovoltaic power generation prediction results.

2. A distributed photovoltaic power generation prediction method based on lightweight feature migration according to claim 1, characterized in that: In step (1), the photovoltaic time series data is decomposed into a low-frequency component representing the main trend and periodic law of photovoltaic output and a high-frequency component representing rapid fluctuations and transient characteristics using discrete wavelet transform.

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

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

5. The distributed photovoltaic power generation prediction method based on lightweight feature migration according to claim 1 is characterized in that: In step (3), the hidden state vector and unit state vector finally obtained by the long short-term memory network of the source domain are used as the initial hidden state vector and unit state vector of the long short-term memory network of the target domain.

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

7. A distributed photovoltaic power generation prediction system based on lightweight feature migration, characterized in that: include: A preprocessing module is 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 length of the source domain low-frequency segment is greater than the length of the other segments; The source domain feature compression and migration module is used to compress the source domain low-frequency segments and convert them into target domain segments. Based on the long short-term memory network, the temporal features of the low-frequency segments and high-frequency segments after compression in the source domain are extracted respectively. And, lightweight migration of high- and low-frequency final state features of the source domain is used to initialize the state vector of the target domain long short-term memory network respectively; The target domain prediction module is used to encode the high-frequency and low-frequency features of the target domain respectively using independent channel long short-term memory networks; 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.

8. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of a distributed photovoltaic power generation prediction method based on lightweight feature migration according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of a distributed photovoltaic power generation prediction method based on lightweight feature migration according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of a distributed photovoltaic power generation prediction method based on lightweight feature migration according to any one of claims 1 to 6 are implemented.

Images