CN-121983962-A - Photovoltaic power prediction method based on full-connection time-space diagram and adaptive attenuation matrix

CN121983962ACN 121983962 ACN121983962 ACN 121983962ACN-121983962-A

Abstract

The invention relates to a photovoltaic power prediction method based on a full-connection time-space diagram and a self-adaptive attenuation matrix, belongs to the technical field of new energy power generation and smart power grids, and particularly relates to multi-station photovoltaic short-term medium-term power prediction. The method solves the problems of too fast information loss at far time, easy disappearance of gradient and poor adaptability of multiple sites caused by the disjoint space-time modeling and attenuation mechanism rigidity of the traditional multivariable time sequence prediction model. The technical scheme includes that the method comprises the steps of multi-source data acquisition and standardized preprocessing, 1D-CNN feature coding integration, full-connection space-time adjacency matrix construction, adaptive attenuation matrix fusion, MPNN space-time feature aggregation and a model of a lightweight MLP prediction layer, and a prediction result is generated through training. The method can remarkably improve the photovoltaic power prediction precision, stability and engineering deployment suitability, and effectively adapt to complex scenes such as mountain full climate and the like.

Inventors

XIANG SHENG
WANG MAOYUAN
LI PENGHUA
HOU JIE
Ilolov Mamadho Ilovic
XIE LIRONG
Jamshed Rachmatov

Assignees

重庆邮电大学

Dates

Publication Date: 20260505
Application Date: 20260128

Claims (10)

1. A photovoltaic power prediction method based on a full-connection time-space diagram and an adaptive attenuation matrix is characterized by comprising the following steps: Firstly, multi-source multi-station data acquisition and standardization preprocessing, wherein numerical weather forecast (Numerical Weather Prediction, NWP) characteristic data and local measurement data (Local Measurement Data, LMD) characteristic data are acquired, the time resolution is unified to be 15 minutes, full-climate operation scenes of a plurality of photovoltaic power stations are covered, cleaning processing is carried out on original data, including filling missing values and removing abnormal values, and a standardized sample tensor is generated by using a sliding window with fixed time length, wherein the shape of the standardized sample tensor is (sample number, time step number, station number and characteristic number), and the data are subjected to standardization processing; Step two, constructing a full-connection space-time diagram prediction model, wherein the model comprises a 1D-CNN feature coding layer, a full-connection space-time adjacency matrix construction layer, an adaptive attenuation matrix fusion module, an MPNN space-time feature aggregation layer and a lightweight MLP prediction layer, and the adaptive attenuation matrix fusion module executes time difference similarity aggregation, adaptive attenuation coefficient calculation and sigmoid constraint attenuation matrix generation; And thirdly, model training and prediction output, wherein a standardized sample is input into the model for training, an Adam optimizer is adopted, data to be predicted is input after training is completed, a photovoltaic power prediction result is generated through model processing, and then the photovoltaic power prediction result is restored into an actual power unit through inverse normalization.
2. The photovoltaic power prediction method based on the full-connection time-space diagram and the adaptive attenuation matrix according to claim 1, wherein in the multi-source multi-station data acquisition and standardization preprocessing step, the cleaning processing comprises filling missing empty values through a block interpolation method, and removing abnormal values beyond a reasonable range, wherein the abnormal values comprise irradiance negative values, power values exceeding the installed capacity of a power station and temperatures exceeding a reasonable interval of-40 ℃ to 60 ℃.
3. The method of claim 1, wherein in the step of multi-source multi-station data acquisition and normalization preprocessing, when the sliding window generates a normalized sample tensor, the time stamps of different station data are forced to be aligned, a four-dimensional tensor format is output as (number of samples, time step, number of stations, feature number), and a Min-Max scaler is fitted separately based on training set data, so that all feature values are mapped to a [0,1] interval.
4. The photovoltaic power prediction method based on the full-connection time-space diagram and the adaptive attenuation matrix according to claim 1, wherein the 1D-CNN feature encoding layer performs convolution operation on the input tensor, and the formula is as follows Wherein Indicating the coding feature of the ith station at the t-th moment, In order for the weights to be trainable, Conv1d is a one-dimensional convolution operation, the input tensor being converted to by dimension adjustment for trainable bias Wherein B is the batch size, L is the fixed window length, N is the number of stations, and D is the feature dimension.
5. The photovoltaic power prediction method based on the full-connection space-time diagram and the adaptive attenuation matrix according to claim 1, wherein the full-connection space-time adjacent matrix construction layer enhances node characteristics through linear mapping, and the formula is Wherein For trainable weights, calculating the similarity between every two nodes through dot products, and constructing a scoring matrix with the formula of Wherein u and v are node indexes, and then the adjacent matrix is obtained through LeakyReLU activation and Softmax normalization Wherein I is an identity matrix.
6. The photovoltaic power prediction method based on the full-connection time-space diagram and the adaptive attenuation matrix according to claim 1, wherein in the adaptive attenuation matrix fusion module, the time difference similarity aggregation aims at time difference Aggregate similarity mean in two cases, v=0 and v > 0; The formula is: When v=0, the number of times of the process, ; When v >0 Where L is the time window size, Is the number of stations.
7. The photovoltaic power prediction method based on the fully connected time-space diagram and the adaptive attenuation matrix according to claim 6, wherein the adaptive attenuation coefficient calculation is characterized in that trainable parameters are calculated through a Sigmoid function Mapping to the (0, 1) interval to obtain Fusing the fixed attenuation reference with the data similarity to obtain a basic attenuation value Then calculate the adaptive attenuation coefficient Where decay is the fixed attenuation reference and the clamp function ensures that the base attenuation value is between [0.3,0.99], k is the super-parameter that controls the overall rate of attenuation.
8. The photovoltaic power prediction method based on the fully connected time-space diagram and the adaptive attenuation matrix according to claim 7, wherein the sigmoid constrained attenuation matrix is generated by the formula Constructing a time decay matrix, wherein i and j are time step indexes, For the time difference, γ is the learning parameter control decay slope, δ is the learning parameter control decay start point, sigmoid is the smooth nonlinear activation function.
9. The method for photovoltaic power prediction based on fully connected space-time diagrams and adaptive attenuation matrices according to claim 1, wherein the MPNN space-time feature aggregation layer captures local space-time dependencies using moving windows, sets window size M=2 and step size s=1, slides along a time axis, each window contains M×N nodes, and uses a message passing phase formula And Aggregating features, and updating the stage formula by the features Updating node characteristics, and finally executing time sequence pooling to extract high-dimensional core characteristics.
10. The photovoltaic power prediction method based on the full-connection time-space diagram and the adaptive attenuation matrix of claim 1, wherein the lightweight MLP prediction layer comprises two layers of perceptrons, the first layer is an active layer formula The second layer is an output layer formula Wherein And In order for the parameters to be trainable, X T sum For trainable parameters, T is the predicted time step, To predict the power matrix, the inverse normalization formula is adopted And restoring to an actual power unit.

Description

Photovoltaic power prediction method based on full-connection time-space diagram and adaptive attenuation matrix Technical Field The invention belongs to the technical field of new energy power generation and intelligent power grids, and relates to a photovoltaic power prediction method based on a full-connection time-space diagram and a self-adaptive attenuation matrix. Background Under the trend of accelerating the transformation of the global energy structure to new energy, the photovoltaic power generation is used as a core component of clean energy, the installed capacity and the grid-connected scale of the photovoltaic power generation continuously climb, and the photovoltaic power generation has a key meaning for improving the new energy consumption level and guaranteeing the safe operation of a power grid. The photovoltaic power output is comprehensively influenced by meteorological factors such as solar irradiance, temperature, wind speed and the like and site equipment characteristics, the characteristic of remarkable fluctuation and space-time coupling is presented, particularly in complex application scenes such as mountain full climate and the like, the meteorological environment is changed severely, the data distribution is unstable, and higher requirements are put on the accuracy, the robustness and the engineering suitability of a prediction model. The existing photovoltaic power prediction method has a plurality of short plates, and is difficult to meet the actual requirements of complex scenes. The traditional multivariable time sequence prediction adopts a two-step method to model the space and time correlation separately, so that the characteristics of coexistence of short-term weather mutation and long-term trend of the photovoltaic scene can not be adapted in a unified graph structure, the coupling relation of the cross-moment and multiple sensors is characterized as insufficient, and the problems of over-fitting or under-fitting are easily caused. Although a part of the full-connection time-space diagram model introduces a time distance attenuation matrix, the general power exponential attenuation mechanism has a rigid defect, so that effective association at a far time can be weakened too quickly, the dependence of photovoltaic power output on weather factors at the far time still exists, and the problem limits the generalization performance of the model. Meanwhile, the problem that the time stamps are difficult to align and the heterogeneity of cross-site equipment is strong exists in the multi-site data, and the traditional preprocessing method cannot effectively weaken interference. For example, when multi-source data is collected, including numerical weather forecast (Numerical Weather Prediction, NWP) feature data and local measurement data (Local Measurement Data, LMD) feature data, the data temporal resolution needs to be uniform, but cross-site heterogeneity leads to alignment difficulties. The existing attenuation matrix often accompanies the problem of gradient disappearance, and further influences model training stability and prediction accuracy. Therefore, the core objective of the invention is to solve the defects of the prior art in aspects of space-time coupling modeling disjoint, unreasonable attenuation logic, poor multi-station adaptability and the like, provide a high-precision prediction method for customizing and optimizing a photovoltaic scene, realize comprehensive capture and effective modeling of space-time dependence and improve the precision, stability and engineering deployment adaptability of photovoltaic power prediction. Disclosure of Invention In view of the above, the invention aims to provide a photovoltaic power prediction method based on a full-connection time-space diagram and an adaptive attenuation matrix, so as to solve the problems of space-time modeling disjoint, attenuation mechanism rigidity, easy disappearance of gradients and poor multi-station suitability of a traditional multivariable time sequence prediction model, and effectively overcome the defect of 'excessively fast attenuation' associated with long-term moments in a general attenuation matrix, thereby remarkably improving the precision, stability and engineering deployment suitability of photovoltaic power prediction. In order to achieve the above purpose, the present invention provides the following technical solutions: A photovoltaic power prediction method based on a full-connection time-space diagram and an adaptive attenuation matrix comprises the following steps: S1, multi-source multi-station point data acquisition and standardized preprocessing. And collecting multi-source operation and meteorological data of the photovoltaic power station, generating a standardized sample by adopting a sliding window with fixed time length after cleaning, filling, outlier rejection and normalization processing, realizing the accurate alignment of the time sequence of the multi-site data