CN-121637430-B - Regional distributed photovoltaic power prediction method and system based on multi-metadata cross-modal fusion

Abstract

The invention relates to a regional distributed photovoltaic power prediction method and system based on multi-element data cross-modal fusion, and belongs to the technical field of photovoltaic power prediction. Firstly, through AG-CNN dynamic perception convolution kernel, implementing adaptive extraction of meteorological data space characteristics and power data space heterogeneity, providing high quality space foundation for cross-modal fusion, secondly, utilizing DR-transducer two-dimensional dynamic attention mechanism to deeply mine cross-modal association of 'history power time sequence-future meteorological drive', adapting fusion requirements under different scenes through dynamic weight distribution, capturing associated evolution in a long period, designing cross-modal characteristic bridging module, and finally, verifying model robustness under typical scenes, and ensuring that the model robustness meets accuracy requirements of multi-scale dispatching of a power grid.

Inventors

• YU YIXIAO
• SUN QIE
• YANG MING
• WANG CHUANQI
• LI MENGLIN

Assignees

• 山东大学

Dates

Publication Date

20260508

Application Date

20260130

Claims (8)

1. 1. The regional distributed photovoltaic power prediction method based on multi-metadata cross-mode fusion is characterized by comprising the following steps of: (1) The space-time data grid consistency processing is carried out, original meteorological data and historical power data are collected, data standardization is carried out, and then space correlation between the meteorological data and the historical power data is achieved by taking uniform geographic coordinates as a reference; Combining coverage areas of the photovoltaic array and weather monitoring points, matching geographic positions of the photovoltaic panels with corresponding weather areas based on spatial scales of weather data, enabling power data and the weather data to form meshed spatial correlation, and simultaneously, carrying out time alignment on the data according to a time axis of predicted granularity, removing abnormal data, and forming a time-space double-dimensional aligned input structure; (2) The AG-CNN carries out self-adaptive extraction on data based on the AG-CNN spatial characteristics of dynamic perception convolution, and the AG-CNN relies on a unified basic spatial reference to realize local modeling through dynamic characteristic perception and cross-region information fusion; Taking grid cells as basic spatial indexes, and real-time adapting the site distribution density and characteristic fluctuation modes in different grid cells by the convolution kernel of AG-CNN through the learnable parameters; the convolution kernel weights of AG-CNN are constructed by the cooperation of a Gaussian function of spatial relative position offset and a learnable parameter: (2) Wherein, (i, j) represents a spatial relative position offset, i is a horizontal spatial relative position offset, j is a vertical spatial relative position offset, σ g is a gaussian kernel bandwidth, s (i, j) is a feature similarity of the offset region and the target region, θ is a learnable parameter, α=5, for balancing the influence of the spatial position relevance and the feature pattern similarity, and when a convolution operation is performed on the feature map, outputting information of a feature adaptive fusion target region and a peripheral association region: (3) Wherein y x,y is a feature value of a target area in the output feature map, and x x+i,y+j is a feature of a peripheral associated area in the input feature map; (3) The cross-modal feature bridging module is designed, and the high-dimensional local features output by AG-CNN are converted into space-time token sequences which can be resolved by DR-transformers through a three-stage processing flow of self-adaptive projection, feature re-parameterization and space-time fusion gating, and key space details and cross-region associated information are reserved; (4) DR-transducer time sequence association modeling integrating double-dimensional dynamic attention is carried out, the DR-transducer takes a space dimension and time sequence dimension dynamic attention mechanism as a core, long time sequence dependence of deep fusion historical power and driving action of future meteorological data are carried out, accurate coupling of meteorological-power cross-mode space-time characteristics is achieved, and a total power prediction result of a distributed photovoltaic system is obtained; DR-transducer realizes cross-modal correlation modeling of historical power-future weather through an encoder-decoder structure; The encoder receives a sequence output by a cross-modal feature bridging module, processes historical power and spatial features, and utilizes a multi-layer stacking structure to introduce a dynamic window division strategy through a bottom layer and strengthen time sequence association of a local region through the calculation of the attention in a window through implicit association and long time sequence dependence among dynamic spatial attention capturing regions; The decoder processes weather data of a future time step, dynamically couples weather features with history features output by an encoder through cross attention under the constraint of a mask self-attention mechanism, adopts a multi-layer stacked structure, and each layer consists of multi-head mask self-attention, multi-head cross attention and a feedforward neural network, wherein the mask self-attention ensures time sequence causality of a weather sequence; And finally, the output layer converts the fused characteristics into regional power predicted values of a plurality of time steps in the future through the full-connection layer, and then the regional power predicted values are weighted and summarized according to the photovoltaic installed capacity of each region to obtain the total power predicted result of the distributed photovoltaic system, so that the full flow from input data to predicted output is completed.
2. 2. The regional distributed photovoltaic power prediction method based on multi-data cross-modal fusion of claim 1, wherein in step (1), the meteorological data comprises solar irradiance, temperature and humidity; The data standardization process is to process meteorological data and historical power data respectively, eliminate the influence of different feature sizes, and adopt a Z-score standardization method, and the formula is as follows: (1) wherein x is the original data, mu is the data mean, sigma is the standard deviation, Is normalized data.
3. 3. The regional distributed photovoltaic power prediction method based on multi-element data cross-modal fusion according to claim 2 is characterized in that a network structure of AG-CNN adopts a multi-layer convolution block series connection design, wherein a first layer is used for capturing instantaneous irradiation mutation in a region through a dynamic kernel, a second layer is used for extracting a cross-regional weather propagation rule through the dynamic kernel, a third layer is used for merging the first two layers of features through convolution to strengthen key details, and each layer of convolution is subjected to batch normalization and activation treatment.
4. 4. The regional distributed photovoltaic power prediction method based on multi-data cross-modal fusion according to claim 3, wherein in the step (3), the feature map F output by AG-CNN is a four-dimensional tensor: (4) Wherein B is the training batch size, C is the characteristic channel number, and H is the number of the corresponding region divisions; firstly, converting a two-dimensional area structure into a one-dimensional sequence through space flattening operation, and realizing the adaptation of channel dimension and transform hidden dimension through learning linear transformation: (5) Wherein, the For flattened sequence features, n=h×w is the total number of regions, In order to project the matrix of the light, D is a transform hidden dimension for the bias term; Because of the difference between the AG-CNN and DR-transducer optimization targets, a re-parameterization layer is introduced to dynamically adjust the feature distribution so as to relieve training offset: (6) Wherein, the In order to gate the coefficients element by element, The function is activated for Sigmoid, 、 The gating mechanism realizes dynamic alignment of the feature distribution of the two models by strengthening key features and suppressing noise for the learnable parameters; to display the temporal feature of the fusion history power and the spatial feature of AG-CNN, a space-time gating unit is introduced, and the associated weights of the temporal feature and the spatial feature are dynamically calibrated through cross-modal attention: Setting historical power sequence Wherein, T h is the historical time step, R is the real number domain, B is the training batch size, N is the total number of regions, D is the hidden dimension of the transducer, and the gating fusion process is as follows: (7) Wherein, the For dynamic fusion coefficients, F in is the output feature of STGU, carrying spatial detail and timing correlation at the same time, STGU is a time-space gating unit for fusing spatial feature and timing feature, W α is a cross-modal attention weight matrix, b α is a cross-modal attention bias term, F reparam is a re-parameterized spatial feature output by equation (6), For a cross-modal attention mechanism, the semantic level calibration of the spatial features and the time sequence features is realized through the association weights of the spatial features and the time sequence features; Final output Directly as input to the DR-transducer encoder.
5. 5. The method for predicting regional distributed photovoltaic power based on multi-data cross-modal fusion of claim 4, wherein DR-transducer focuses on a dynamic attention mechanism of space dimension and time sequence dimension, namely, in the space dimension, attention weight is self-adaptively adjusted through regional feature similarity and relative position relation; Aiming at the discrete distribution characteristics of the distributed photovoltaic sites, the encoder of the DR-transducer calculates the attention weight by implicit association among dynamic space attention capturing areas while considering the similarity of the relative space distance and the characteristic mode, and the calculation formula of the dynamic space attention weight is as follows: (8) wherein q, k, v are query, key, value vector, d k is feature dimension, Beta (r) is a relative spatial distance coefficient for the feature similarity coefficient, and through cooperative modulation of the two, the attention weight can embody natural continuity in space and can focus on a related area with similar features in a targeted manner; The decoder of DR-transducer fuses future meteorological data with the history feature dynamic output by the encoder, and cross-modal association is realized through masking self-attention and cross-attention; masking self-attention forces the model to follow time sequence causality through a lower triangular matrix, and only allows the current time step to pay attention to historical and contemporaneous meteorological features; let the future weather sequence input by the decoder be T is the number of future prediction steps, D is the dimension of the meteorological feature vector, R is the real number domain, m t is the meteorological feature vector of the T step, and the query is generated through linear projection Key and key Value of , , , Is a meteorological characteristic projection matrix, h is the number of attention heads; for the t-th time step, the similarity of the masked self-attention is calculated as: (9) Wherein, the In order for the time step to be of interest, Is a scaling factor; introducing a lower triangular mask matrix Future information is suppressed by masking: (10) In the formula, Representing the t-th row in the mask matrix The elements of the column, MASKEDSIM, are the masked attention similarity scores; Final first The masked self-attention output MASKEDATTN (T) for the step is: (11) Wherein, the For outputting a projection matrix, v represents a value vector, softmax (·) is a normalized exponential function, concat (·) represents multi-head attention splicing operation, and a multi-head attention result is spliced and projected to obtain meteorological features containing time sequence dependence 。
6. 6. The method for predicting regional distributed photovoltaic power based on multi-data cross-modal fusion of claim 5, wherein a multi-objective loss function oriented to regional characteristics and cross-modal constraints is constructed in the method.
7. 7. The regional distributed photovoltaic power prediction method based on multi-data cross-modal fusion according to claim 6, wherein the multi-objective loss function adopts a mean square error as a basic loss term, and the formula is as follows: (12) Wherein N is the total number of the areas, In order to predict the step of time, And The predicted power and the real power of the ith area at the moment t are respectively; AG-CNN extracts local features through dynamic kernel and cross-regional fusion, but the regional features output by AG-CNN cause adjacent regional association fracture due to the randomness of dynamic neighborhood, the core of regional feature consistency loss is to force a model to respect the implicit association among regions in a physical space, so that feature extraction is ensured to be focused on the local details of a single region, and the cooperative change of adjacent regions due to common meteorological disturbance can be captured: (13) Wherein N i is a dynamic neighborhood of the region i, w i,j is a feature similarity weight of the region i, j, and the higher the similarity is, the larger the weight is; the core of DR-transducer is to merge historical power time sequence characteristics and future weather driving characteristics through cross attention, the goal of cross-modal time sequence association loss is to force a model to learn the physical causal relationship between weather factors and power changes, ensure that cross-modal association in the time sequence dimension accords with objective rules, and not just fit data surface correlation, and through double constraint, the cross-modal association in the time sequence dimension is learned from association of a physical mechanism layer specification model: (14) Wherein, the As a trend consistency loss function to constrain the physical rationality of the direction of change, β is a weight coefficient of the trend consistency loss for balancing the strength of the trend constraint with the attention weight constraint: (15) Wherein, the In order to predict the amount of time-series variation in power, Is the time sequence variation of the meteorological factors, For cross entropy loss, high punishment is given to the situation that the power trend is opposite to the meteorological trend; objective relevance to constrain attention weights as a cross-modal attention loss function: (16) Wherein, the In DR-transducer cross attention, meteorological features are specific to regions At the position of The attention weight of the moment in time, Weather-power true correlation coefficients calculated for historical data; the overall loss function is a weighted sum of the three losses described above: (17)。
8. 8. The regional distributed photovoltaic power prediction system based on the multi-metadata cross-modal fusion is applied to the regional distributed photovoltaic power prediction method based on the multi-metadata cross-modal fusion as claimed in claim 1, and is characterized by comprising the following steps: The data processing module is used for carrying out space-time data meshing consistency processing, collecting original meteorological data and historical power data, carrying out data standardization, and then realizing space correlation between the meteorological data and the historical power data by taking uniform geographic coordinates as a reference; the extraction module is used for carrying out self-adaptive extraction on data based on AG-CNN spatial features of dynamic perception convolution, and AG-CNN is based on a unified basic spatial reference, and local modeling is realized through dynamic feature perception and cross-region information fusion; The design module is used for designing a cross-modal feature bridging module, converting the high-dimensional local features output by AG-CNN into a space-time token sequence which can be resolved by DR-transducer through a three-stage processing flow of self-adaptive projection, feature re-parameterization and space-time fusion gating, and simultaneously reserving key space details and cross-region association information; The prediction module is used for integrating DR-transducer time sequence association modeling of the two-dimensional dynamic attention, the DR-transducer takes a space dimension and time sequence dimension dynamic attention mechanism as a core, long time sequence dependence of deep fusion historical power and driving action of future meteorological data are achieved, accurate coupling of meteorological-power cross-mode space-time characteristics is achieved, and a total power prediction result of the distributed photovoltaic system is obtained.

Description

Regional distributed photovoltaic power prediction method and system based on multi-metadata cross-modal fusion Technical Field The invention relates to a regional distributed photovoltaic power prediction method and system based on multi-element data cross-modal fusion, and belongs to the technical field of photovoltaic power prediction. Background The current distributed photovoltaic power prediction research forms a multi-technology path, but the prior method has double limitations in the aspects of power-weather cross-modal fusion and long time sequence dependent modeling, and always restricts the improvement of prediction precision. From the physical model, the photovoltaic module conversion mechanism-based photovoltaic module conversion mechanism construction can embody the direct influence of meteorological factors on power and show a certain reliability in short-term prediction, but the model depends on fixed module parameters and high-resolution meteorological data, the method is difficult to describe long-period accumulation effects of meteorological factors (such as cloud cover and temperature), lacks dynamic cross-modal adaptation of power and meteorological data, and cannot cope with fusion requirements under meteorological mutation or data heterogeneous scenes only by simply correlating the two through a mechanism formula. Limitations of statistical models are focused on the lack of shallow and nonlinear adaptation across modality associations. The model establishes association through mining historical data rules, basic prediction can be realized when data distribution is stable, but not only nonlinear relation between weather and power cannot be captured, but also cross-modal information utilization is limited to shallow statistical association (such as simple linear regression), when modal difference between weather data and power data is increased (such as power fluctuation aggravated by irradiation dip in overcast and rainy weather), association description of the model on weather driving-power response is rapidly distorted, and accuracy is rapidly reduced. These limitations do not exist in isolation, but are interleaved with each other around two major core problems of cross-modal fusion and long time sequence dependence, and the two core problems are progressive layer by layer, so that modeling difficulty of distributed photovoltaic power prediction is formed together, namely, whether weather and power data can be effectively cooperated is determined, whether the cooperated relationship can be stably transferred in the time dimension is influenced, and modeling complexity is further amplified by coupling the two. The above challenges are superimposed and ultimately point to the top-level architecture bottleneck of insufficient sequential-spatial-cross-modal collaborative modeling capability. On one hand, the multi-model splitting space association and the cross-mode fusion are realized, for example, partial space-time models capture site space association only at a power level and do not take weather space distribution into cooperative consideration, on the other hand, the traditional time sequence architecture is difficult to bear long-period cross-mode association transfer, an LSTM memory unit is attenuated along with time, early key weather features are easy to lose, attention is easy to disperse under a long sequence of a transducer, and a core period of weather trend-power trend is difficult to focus. The synergetic absence makes the existing model difficult to consider the regional difference and long-term coupling, and severely restricts the generalization capability and the robustness of the prediction model under the complex meteorological conditions. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a regional distributed photovoltaic power prediction method and system based on multi-metadata cross-modal fusion, which take distributed photovoltaic accurate power prediction as a core target, focus power-weather cross-modal deep fusion and long time sequence dependence accurate capture collaborative breakthrough, combine the technical characteristics of self-adaptive spatial convolution (AG-CNN) and dynamic associated transformers (DR-transformers), construct an end-to-end combined prediction framework, firstly realize adaptive extraction of weather data spatial characteristics and power data spatial heterogeneity through a dynamic perception convolution kernel of AG-CNN, not only keep continuous structures of weather data, but also attach the characteristic of distributed photovoltaic 'point multi-aspect wide', provide a high quality spatial basis for cross-modal fusion, secondly utilize a two-dimensional dynamic attention mechanism of DR-transformers, deeply mine cross-modal association of 'history power time sequence-future weather driving', simultaneously capture the association in a long period through the different scenes by dynamic wei