Cross-modal net load prediction method based on public weather forecast

By adopting cross-modal comparison learning and attention mechanisms in net load prediction, public meteorological forecasting and timing meteorological data are integrated, which solves the problem that the impact of meteorological factors in the existing technology is difficult to accurately predict, and efficient net load prediction is achieved. Reduced costs.

CN118940884APending Publication Date: 2024-11-12NORTH CHINA ELECTRIC POWER UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410968528.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

When the existing net load prediction methods consider meteorological factors, especially after the penetration rate of distributed photovoltaic systems increases, it is difficult to accurately predict the previous net load, and the cost of obtaining numerical meteorological forecast data is high, which affects the prediction accuracy.

Method used

Using a cross-modal net load prediction method based on public meteorological forecasting, the feature extraction and fusion of public meteorological forecasts and timing meteorological data through cross-modal comparison learning and attention mechanisms is adopted to enhance the prediction effect, and avoid logarithmic meteorological forecasting. rely.

Benefits of technology

It improves the accuracy of net load prediction, reduces the prediction cost, and enhances the characteristic representation effect of public meteorological forecasts, which are suitable for situations where numerical meteorological forecasts cannot be obtained.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940884A_ABST
    Figure CN118940884A_ABST
Patent Text Reader

Abstract

The invention provides a cross-modal net load prediction method based on public weather forecast, and belongs to the technical field of power system net load prediction. Public weather forecast data is introduced to serve as day-ahead weather information, weather feature semantics are enriched in a cross-modal comparative learning mode, then historical net load features and weather features are fused in a cross-modal fusion mode, and the fused features are regressed to obtain a day-ahead prediction result. According to the invention, effective embedding of public weather forecast information is realized, and a feature representation effect is enhanced for a downstream prediction task; the adopted cross-modal attention mechanism can be used for fusing historical net load features and public weather forecast features, so that the influence of isomerism between the two features is reduced, and the feature fusion effect is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention provides a cross-modal net load prediction method based on public weather forecast, belonging to the technical field of power system net load prediction. Background Art

[0002] As the global demand for sustainable energy solutions continues to increase, the penetration rate of renewable energy generation in the power grid continues to increase, and its contribution to global electricity supply is increasing. Photovoltaic systems are considered to have the greatest potential for emission reduction, among which small distributed generation systems will dominate the future growth of photovoltaic installed capacity. However, these distributed photovoltaic systems are mostly installed on the user side due to costs, policies, etc., and the "behind-the-meter photovoltaic" problem has led to changes in the operation and planning of the power system, especially in load forecasting during critical periods of the day. The high penetration rate of behind-the-meter photovoltaics increases the peak-to-valley difference of the load. The load will drop significantly at the peak of photovoltaic power generation at noon, resulting in a duck-shaped net load curve, which is called the duck curve problem, and poses a huge challenge to the supply and demand balance of the power grid.

[0003] Among the existing net load forecasting methods, some methods consider using historical power data to extrapolate the day-ahead net load, or consider the impact of meteorological factors on the net load during the extrapolation, and use historical meteorological measured data as covariates to improve the prediction accuracy. With the development of weather forecasting technology, numerical weather forecasts have become a feasible covariate for net load forecasting tasks, which can effectively improve the prediction accuracy. However, in practical applications, the acquisition of numerical weather forecasts usually requires cooperation with data providers or meteorological stations in the form of organizations or enterprises, which often involves a large amount of service fees, making it challenging to obtain specific numerical weather forecast data for the corresponding geographical area.

[0004] In general, there are some challenges in the application of day-ahead meteorological information in net load forecasting.

[0005] One of the existing technologies, aimed at the net load power uncertainty problem, mainly considers the periodicity and regularity of net load power fluctuation, and extrapolates the day-ahead power through the historical net load power. The specific form is as follows:

[0006] First, select a suitable prediction network, such as the Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), Temporal Convolutional Network (TCN) and other deep learning methods, and then design the corresponding network structure and network input features according to the characteristics of net load power uncertainty. Secondly, the network learns the historical net load power fluctuation characteristics and adjusts the network parameters through the loss function to improve the final prediction effect of the network. Finally, use the trained prediction network model to perform day-ahead net load prediction.

[0007] This technology does not consider the impact of distributed photovoltaics on net load from a meteorological perspective. As the penetration rate of distributed photovoltaics in net load increases, the impact of meteorological factors on net load increases. In the prior art 1, since meteorological factors are not considered, the prediction results are poor in samples with significant fluctuations in meteorological conditions, and the trend and amplitude of the predicted curve are significantly different from the actual curve.

[0008] The second existing technology considers the impact of weather on net load power and uses historical measured weather data or day-ahead numerical weather forecast data as forecast task covariates to improve the forecast accuracy of day-ahead net load. The specific form is as follows:

[0009] First, select a suitable prediction network, such as the Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), Temporal Convolutional Network (TCN) in deep learning methods, and design the corresponding network structure and network input features. Secondly, splice the historical net load power with the historical measured meteorological data or the day-ahead numerical meteorological forecast data and input them into the network. The network learns the net load characteristics and meteorological characteristics, and adjusts the network parameters through the loss function to improve the final prediction effect of the network. Finally, use the trained prediction network model, input the corresponding historical net load data and meteorological data, and predict the day-ahead net load power.

[0010] The method of using historical measured meteorological data lacks a description of the meteorological conditions on the previous day, and the improvement in prediction effect is limited; the method of using numerical meteorological forecast data on the previous day can effectively improve the accuracy, but in practical applications, the acquisition of numerical meteorological forecasts usually requires cooperation with data providers or meteorological stations in the form of organizations or enterprises, which often involves a large amount of service fees, making the acquisition of numerical meteorological forecast data in specific areas challenging and hindering the realization of prediction tasks. Summary of the invention

[0011] In response to the shortcomings of one of the technologies, the present invention considers historical net load power and public weather forecast data in the prediction task, wherein the public weather forecast can provide day-ahead weather information, thereby ensuring the lower limit of the net load prediction effect in the absence of numerical weather forecasts.

[0012] In view of the shortcomings of the second technique, the present invention considers public weather forecast as the day-ahead weather information in the net load forecasting task. Compared with public weather forecast, residents usually cannot obtain numerical weather forecast information, and its impact on residents' behavior is minimal; at the same time, public weather forecast is open to the public free of charge and has a potential impact on residents' travel plans. Therefore, it has a certain reference value for net load forecasting and can be used as day-ahead weather data when numerical weather forecast is not available.

[0013] In response to the problem that public weather forecasts have less information than numerical weather forecasts, the present invention may consider using cross-modal learning technology to narrow the distance between public weather forecasts and information-rich time-series meteorological data in the feature space, thereby enhancing the feature representation effect in the public weather forecast, increasing the amount of information, and improving the effectiveness of downstream prediction tasks.

[0014] In order to solve the heterogeneity problem between historical net load information and day-ahead meteorological information, the present invention introduces a cross-modal attention mechanism to dynamically allocate weights to the feature encodings of the two, thereby improving the fusion effect of the two information features.

[0015] The specific technical solution of the present invention is:

[0016] A cross-modal net load prediction method based on public weather forecast introduces public weather forecast data as the day-ahead meteorological information, enriches its meteorological feature semantics through cross-modal comparative learning, and then uses cross-modal fusion to fuse historical net load features and meteorological features, and regresses the fused features to obtain the day-ahead prediction results.

[0017] The specific steps include:

[0018] S1. Feature extraction is performed on historical net load data and public weather forecasts respectively.

[0019] Adopt meteorological feature coding module, including feature coding and comparative learning;

[0020] First, the corresponding feature extraction encoders are used to map different meteorological data into feature vectors. Then, contrastive learning is used to calculate the contrastive loss between different feature vectors to optimize the network parameters of the feature encoder.

[0021] Among them, the contrastive learning loss function uses the optimal mutual information method to optimize the parameters of the feature extraction module. The calculation formula is as follows:

[0022]

[0023] Where: is the i-th data sample pair, sim represents the similarity function; τ is the learnable parameter in contrastive learning; L p is the training loss of the iterative public weather forecast encoder, L m is the training loss of the symmetrical temporal meteorological encoder, and the final total loss function is L = (L p +L m ) / 2.

[0024] The structure of the feature encoder is designed according to the formal characteristics of different meteorological information, namely the public weather forecast encoder and the time series meteorological encoder.

[0025] Public weather forecasts can provide discrete features of the day's weather. Therefore, the feature encoder constructs multiple linear projection layers in a layer-by-layer manner and skips the input and output through residuals to avoid information loss.

[0026] Time series meteorological data can provide the network with time series information of multi-dimensional variables, so the time series meteorological encoder introduces a time series convolutional network for feature extraction. Depth-wise separable convolution is introduced in the time series meteorological encoder. It consists of depth-wise convolution and point-wise convolution. In depth-wise convolution, a large convolution kernel is used to increase the effective receptive field. At the same time, each convolution kernel only processes one feature channel to facilitate the extraction of local patterns in the time series. After depth-wise convolution, the correlation features between different feature channels are combined with the feature maps of all channels through point-wise convolution to generate a new feature map to extract the correlation dependencies between meteorological features.

[0027] S2. Extract the time series features of the historical power sequence and introduce a time series encoder composed of a temporal convolutional network (TCN) to encode the historical net load information. The encoder maps the historical net load sequence code to the latent space, and then regresses the linear projection layer to obtain the day-ahead net load forecast value.

[0028] S3, the regression stage is based on the encoder trained in the pre-training stage, and further constructs a feature fusion module, which is composed of a cross-modal attention mechanism. The feature fusion module accepts two types of feature encoding inputs, namely historical information encoding and and meteorological information codes

[0029] First, the mapping representation of the two features is obtained through the linear projection layer, and then the corresponding elements of the features are added; the historical net load feature vector is taken as the main factor, and the two features are combined to form the result. The weights of the historical net load feature vectors are assigned under the Softmax function, and finally the residual operation is performed with the embedded historical net load feature vectors to obtain the fusion features, and then the day-ahead net load forecast is obtained through the linear regression layer. The relevant formula is as follows:

[0030] X q =(W h v h +b h )+(W m v m +b m ) (2)

[0031]

[0032] X o =(W e X a +b e )+(W v X h +b v )(4)

[0033] Where W h , W m , W v , W e is the weight of the linear projection layer, b h 、b m 、b v 、b e for deviation; Represents the multiplication of elements at corresponding positions between two matrices.

[0034] The technical effects of the present invention are:

[0035] 1) Public weather forecasts are free and open to the public, which effectively avoids the high cost of using commercial numerical weather forecast data. In addition, public weather forecasts are widely followed by residents and have a potential impact on their electricity usage habits. Using public weather forecasts as covariates for day-ahead net load forecasts in the absence of numerical weather forecasts can help improve the accuracy of net load forecasts.

[0036] 2) The cross-modal contrastive learning module can enrich the semantic features of public weather forecasts. This module is used to capture the latent spatial feature relationship between the discrete information of public weather forecasts and the time series meteorological data, realize the effective embedding of public weather forecast information, and enhance the feature representation effect for downstream prediction tasks.

[0037] 3) The cross-modal attention mechanism adopted by the cross-modal feature fusion module can be used to fuse historical net load features and public weather forecast features, thereby reducing the impact of the heterogeneity between the two features and improving the effect of feature fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a model prediction flow chart of the present invention;

[0039] Figure 2 It is a flow chart of the meteorological feature extraction module of the present invention;

[0040] Figure 3 It is a structural diagram of a time series meteorological data encoder of the present invention;

[0041] Figure 4 This is a structural diagram of the cross-modal feature fusion module of the present invention. DETAILED DESCRIPTION

[0042] The specific technical solution of the present invention is described in conjunction with the accompanying drawings.

[0043] The cross-modal net load prediction method based on public weather forecast is as follows: Figure 1 shown.

[0044] S1. Feature extraction is performed on historical net load data and public weather forecasts respectively.

[0045] The meteorological feature coding module includes two parts: feature coding and comparative learning.

[0046] First, the corresponding feature extraction encoders are used to map the time series meteorological data and public weather forecast data into feature vectors. Then, contrastive learning is used to calculate the contrast loss between different feature vectors to optimize the network parameters of the feature encoder. The specific process of the meteorological feature encoding module is as follows: Figure 2 shown.

[0047] Constructing positive and negative sample pairs is an important part of contrastive learning. It is similar to image and text sample pairs. Public weather forecasts and time series weather data of the same day naturally form positive sample pairs. On the contrary, time series weather data of different dates have low similarity, so negative sample pairs are formed. The contrastive learning loss function uses the optimal mutual information method to optimize the parameters of the feature extraction module. The calculation formula is as follows:

[0048]

[0049] Where: is the i-th data sample pair, sim represents the similarity function; τ is the learnable parameter in contrastive learning; L p is the training loss of the iterative public weather forecast encoder, L mis the training loss of the symmetrical temporal meteorological data encoder, and the final total loss function is L = (L p +L m ) / 2.

[0050] Since there are formal differences between the public weather forecast features and the time series weather data features, the structure of the feature encoder is designed according to the formal characteristics of different weather information, namely the public weather forecast encoder and the time series weather encoder.

[0051] Public weather forecasts usually provide a rough description of the weather the day before, such as weather type, wind speed level, average temperature and other discrete quantities, which are suitable for mapping using linear projection layers. In order to enhance the robustness of feature extraction operations, the discrete feature encoder constructs multiple linear projection layers in a layer-by-layer manner, and skips the input and output through the residual method to avoid information loss.

[0052] Time series meteorological data can provide the network with time series information of multi-dimensional variables. Compared with discrete quantities, there is a time series correlation between different time sections, which can reflect the detailed fluctuations of daily weather. In order to capture the complex patterns between multi-dimensional time series data, the encoder introduces a time series convolutional network for feature extraction. In order to further increase the effective receptive field, improve the information capture range of convolution, and extract the dependency correlation between multiple variables, the encoder introduces a depth-wise separable convolution. It consists of depth-wise convolution and point-wise convolution. Compared with ordinary convolution, the depth-wise convolution uses a large convolution kernel to increase the effective receptive field. At the same time, each convolution kernel only processes one feature channel, thereby ensuring that the feature channels will not be mixed, so as to facilitate the extraction of local patterns in the time series. After the depth-wise convolution, the correlation features between different feature channels need to be combined through point-wise convolution to merge the feature maps of all channels, thereby generating a new feature map to extract the correlation dependencies between meteorological features. The general structure of the time series meteorological data encoder is as follows: Figure 3 shown.

[0053] S2. The historical net load power sequence contains rich time series information, such as trends, seasonality, and periodicity, which can provide certain characteristic information during prediction. In order to better extract the time series characteristics of the historical power sequence, a time series encoder composed of a temporal convolutional network (TCN) is also introduced to encode the historical net load information. The encoder maps the historical net load sequence code to the latent space, and then regresses the linear projection layer to obtain the day-ahead net load forecast value.

[0054] S3, the regression stage is based on the encoder trained in the pre-training stage to further build a prediction model. This module is composed of a cross-modal attention mechanism. The general structure is as follows Figure 4This module accepts two types of feature encoding inputs, namely historical information encoding and meteorological information codes First, the mapping representation of the two features is obtained through the linear projection layer, and then the corresponding elements of the features are added; the historical net load feature vector is taken as the main factor, and the two features are combined to form the result. The weights of the historical net load feature vectors are assigned under the Softmax function, and finally the residual operation is performed with the embedded historical net load feature vectors to obtain the fusion features, and then the day-ahead net load forecast is obtained through the linear regression layer. This fusion method enables the meteorological information features to be complementary to the historical information features, supplementing the semantics of the meteorological information features, thereby expressing richer information. The relevant formula is as follows:

[0055] X q =(W h v h +b h )+(W m v m +b m ) (2)

[0056]

[0057] X o =(W e X a +b e )+(W v X h +b v ) (4)

[0058] Where W h , W m , W v , W e is the weight of the linear projection layer, b h 、b m 、b v 、b e for deviation; Represents the multiplication of elements at corresponding positions between two matrices.

[0059] The present invention introduces a net load prediction model based on public weather forecasts to ensure the prediction effect of the model when there is no numerical weather forecast due to high service fees or other reasons. In view of the problem of limited information content of public weather forecasts, the meteorological feature extraction module introduces a cross-modal contrastive learning method, which uses a feature encoder to map public weather forecasts and time-series meteorological data into a latent space, narrowing the feature distribution gap between the two, thereby enriching the semantic features of the public weather forecast. Taking into account the heterogeneity between public weather forecasts and historical net load data, the fusion module adopts a cross-modal attention mechanism to dynamically allocate the weights of historical net load features and public weather forecast features to improve the feature fusion effect.

Claims

1. A cross-modal net load prediction method based on public weather forecast, characterized in that: Public weather forecast data is introduced as the day-ahead weather information, and its meteorological feature semantics is enriched through cross-modal comparative learning. Then, the cross-modal fusion method is adopted to fuse the historical net load characteristics and meteorological characteristics, and the fused characteristics are regressed to obtain the day-ahead forecast results.

2. The cross-modal net load prediction method based on public weather forecast according to claim 1 is characterized in that: The specific steps include: S1. Extract features from public weather forecasts; Adopt meteorological feature coding module, including feature coding and comparative learning; Firstly, the corresponding feature extraction encoder is used to map different meteorological data into feature vectors. Then, contrastive learning is used to calculate the contrast loss between different feature vectors to optimize the network parameters of the feature encoder. The structure of the feature encoder is designed according to the form characteristics of different meteorological information, namely, the public meteorological forecast encoder and the time series meteorological encoder. The public meteorological forecast can provide discrete features of the day-ahead weather, so the feature encoder adopts a layer-by-layer increase to construct a multi-layer linear projection layer, and jumps the input and output through the residual method to avoid information loss. The time series meteorological data can provide the network with time series information of multi-dimensional variables, so the time series meteorological encoder introduces a time series convolutional network for feature extraction. S2. Extract the time series features of the historical power sequence, introduce a time series encoder composed of a time series convolutional network (TCN), and encode the historical net load information. When training the encoder, the encoder maps the historical net load sequence encoding to the latent space, and then regresses the day-ahead net load forecast value through a linear projection layer, and calculates the loss to optimize the encoder network parameters. S3, the regression stage is based on the encoder trained in the pre-training stage, and further constructs a feature fusion module. The feature fusion module is composed of a cross-modal attention mechanism; the feature fusion module accepts two feature encoding inputs, namely historical information encoding and meteorological information codes First, the mapping representation of the two features is obtained through the linear projection layer, and then the corresponding elements of the features are added; the historical net load feature vector is taken as the main factor, and the two features are combined to form the result. The historical net load feature vector is weighted under the action of the Softmax function, and finally a residual operation is performed with the embedded historical net load feature vector to obtain the fusion feature, and then the day-ahead net load forecast is obtained through the linear regression layer.

3. The cross-modal net load prediction method based on public weather forecast according to claim 2 is characterized in that: In S1, the contrastive learning loss function of contrastive learning uses the optimal mutual information method to optimize the parameters of the feature extraction module. The calculation formula is as follows: Where: is the i-th data sample pair, sim represents the similarity function; τ is the learnable parameter in contrastive learning; L p is the training loss of the iterative public weather forecast encoder, L m is the training loss of the symmetrical temporal meteorological encoder, and the final total loss function is L = (L p +L m ) / 2; Depthwise separable convolution is introduced in the time series meteorological encoder, which consists of depthwise convolution and pointwise convolution. Large convolution kernels are used in depthwise convolution to increase the effective receptive field. At the same time, each convolution kernel only processes one feature channel to facilitate the extraction of local patterns in the time series. After depthwise convolution, the correlation features between different feature channels are combined with the feature maps of all channels through pointwise convolution to generate a new feature map to extract the correlation dependencies between meteorological features.

4. The cross-modal net load prediction method based on public weather forecast according to claim 2 is characterized in that: In S3, the formula related to the fusion feature is as follows: X q =(W h v h +b h )+(W m v m +b m ) (2) X o =(W e X a +b e )+(W v X h +b v ) (4) Where W h , W m , W v , W e is the weight of the linear projection layer, b h 、b m 、b v 、b e for deviation; Represents the multiplication of elements at corresponding positions between two matrices.

Citation Information

Cited By

  • Power load prediction method and device based on causal characteristics, and medium

    CN119315550A