Regional ultra-short-term new energy power prediction method based on multi-modal information fusion

Abstract

The invention belongs to the technical field of power prediction. The invention provides a regional ultra-short-term new energy power prediction method based on multi-modal information fusion, and the method comprises the steps: obtaining satellite remote sensing data, station observation data, meteorological mode data and station attribute data of a target region, and carrying out the time-space alignment and abnormal value processing; performing feature extraction from the data; performing multi-level fusion on the extracted features by using a feature fusion network to obtain fused features; and inputting the fusion features into a prediction model to obtain the predicted ultra-short-term new energy power of the target area. According to the method, multi-modal data features are integrated by using a multi-level fusion algorithm, more comprehensive and targeted data support is provided for regional ultra-short-term new energy power prediction, the performance and generalization ability of a prediction model can be significantly improved, a power dispatching department can make a decision more accurately, and the prediction efficiency is improved. And efficient consumption of new energy and stable operation of a power grid are realized.

Description

Technical Field

[0001] The present invention relates to the field of power prediction technology, and in particular to a method for regional ultra-short-term new energy power prediction based on multimodal information fusion. Background Art

[0002] Amidst the global energy transition, renewable energy generation, as a core component of clean energy, continues to see increasing penetration in power systems. For provincial-level regions, the large-scale, centralized integration of renewable energy into the grid poses significant challenges to the safe and stable operation of the power system. Ultra-short-term renewable energy generation output forecasts (for the next 0-4 hours) are crucial for power dispatch decisions, and their accuracy is directly impacted by grid frequency and peak regulation, power market transactions, and the efficiency of renewable energy consumption.

[0003] The provincial region is vast, with complex and diverse topography, encompassing different geographical units such as plains, mountains, basins, and coastal areas, and meteorological conditions show significant spatial heterogeneity. The output of renewable energy power generation (especially photovoltaic and wind power) is highly dependent on meteorological factors such as light intensity, wind speed, temperature, and cloud cover, and the temporal and spatial distribution of these factors in different regions of the province vary greatly. For example, topographic obstruction in mountainous areas can cause sudden changes in local wind speed, sea-land breeze circulation in coastal areas can cause fluctuations in wind power generation output, and the rapid migration of clouds can cause dramatic fluctuations in photovoltaic power generation output. This strong spatiotemporal variability makes it difficult for a single data source or simple model to accurately capture the dynamic changes in the province's renewable energy power generation output.

[0004] Currently, provincial-level ultra-short-term renewable energy power generation output forecasts rely primarily on three types of data: first, site observation data, including historical renewable energy site power generation output and wind speed and irradiance data collected by on-site meteorological stations; second, meteorological model data, namely gridded meteorological elements output by numerical weather forecasts (NWP); and third, satellite remote sensing data, such as satellite cloud maps and surface albedo from MODIS. However, existing technologies have significant limitations in data utilization and integration, primarily in the following aspects: From the perspective of data sources, while field observation data can reflect detailed information at a single point, it is limited by the spatial distribution density of observation sites, making it difficult to cover all new energy-rich areas in the province. Furthermore, it is susceptible to equipment failures, transmission delays, and other factors, resulting in insufficient data integrity and timeliness. While meteorological model data has province-wide spatial coverage, its spatial resolution is relatively low (typically 3-10 kilometers), making it difficult to depict local meteorological anomalies caused by micro-topography such as valleys and building complexes. Furthermore, forecast errors accumulate as the forecast period lengthens. Satellite remote sensing data can provide macroscopic information such as cloud movement and surface radiation over a wide range of areas, but suffers from low temporal resolution (e.g., updates every 15-30 minutes) and interference from cloud and rain obstructions, making it difficult to meet the high-frequency dynamic monitoring needs of ultra-short-term forecasts.

[0005] From the perspective of fusion methods, traditional forecasting models often use simple splicing or weighted overlay to process multi-source data, failing to deeply explore the inherent connections between different modal data. For example, machine learning-based models are often trained solely using historical station data, ignoring future irradiance trends implied by satellite cloud images. While physical models incorporate meteorological model data, they fail to effectively utilize station observation data for real-time correction of forecast deviations. This "data island" phenomenon prevents the full complementarity of multi-source information. Under complex meteorological conditions (such as severe convective weather and rapid temperature changes), forecast errors often significantly exceed the acceptable range for actual scheduling.

[0006] Furthermore, provincial-level renewable energy sites are diverse (centralized wind power, distributed photovoltaics, and agricultural photovoltaics), with significantly different power generation characteristics across different types of sites, often influenced by non-meteorological factors such as equipment aging and maintenance status. Existing forecasting methods often focus on a single energy type or homogeneous region, lacking the ability to collaboratively model diverse renewable energy sites across the province. This makes it difficult to adapt to the provincial power grid's forecasting needs for the total renewable energy generation output across the region.

[0007] Therefore, how to break through the limitations of single-source data, deeply integrate satellite remote sensing, station observation and meteorological model data through multimodal fusion algorithms, fully tap the complementary value of different data in time and space scales and information dimensions, and improve the accuracy and robustness of the province's ultra-short-term new energy power generation output forecast is a key issue that needs to be urgently solved in the field of new energy power systems. Summary of the Invention

[0008] In order to solve the technical problems existing in the above background technology, the present invention provides a regional ultra-short-term new energy power prediction method based on multimodal information fusion.

[0009] The present invention provides a method for regional ultra-short-term new energy power prediction based on multimodal information fusion, comprising the following steps: Obtain satellite remote sensing data, station observation data, meteorological model data, and station attribute data of the target area, and perform spatiotemporal alignment and outlier processing; Extract cloud system dynamic characteristics and surface radiation characteristics from satellite remote sensing data, extract power time series characteristics and local meteorological fluctuation characteristics from station observation data, extract regional meteorological trend characteristics from meteorological model data, and extract station type characteristics and equipment status characteristics from station attribute data. A feature fusion network is used to perform multi-level fusion of the extracted cloud system dynamic features, surface radiation features, power time series features, local meteorological fluctuation features, regional meteorological trend features, station type features, and equipment status features to obtain fused features. The number of fusion levels is determined at least based on the data feature complexity index, the prediction scenario dynamic index, and the feature correlation strength index. The fusion features are input into a prediction model to obtain the predicted ultra-short-term new energy power of the target area.

[0010] The present invention uses a multi-level fusion algorithm to integrate multimodal data features to provide more comprehensive and targeted data support for regional ultra-short-term renewable energy power forecasts, which can significantly improve the performance and generalization ability of the prediction model, help power dispatching departments make more accurate decisions, and achieve efficient consumption of renewable energy and stable operation of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a flow chart of a method for regional ultra-short-term renewable energy power prediction based on multimodal information fusion disclosed in an embodiment of the present invention; Figure 2 It is a scene schematic diagram of the solution of the present invention; Figure 3 is a schematic diagram of a process for determining the number of fusion levels disclosed in an embodiment of the present invention; Figure 4 is a schematic diagram of a process for obtaining fusion features through multi-level fusion disclosed in an embodiment of the present invention; Figure 5 This is a flow chart of predicting the ultra-short-term new energy power of a target area based on fusion features disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0013] like Figure 1 、 Figure 2 As shown, the embodiment of the present invention discloses a method for regional ultra-short-term new energy power prediction based on multimodal information fusion, comprising the following steps: S10, obtaining satellite remote sensing data, station observation data, meteorological model data and station attribute data of the target area, and performing spatiotemporal alignment and outlier processing.

[0014] For target areas (e.g., provincial-level divisions), the three key data types mentioned above are acquired simultaneously. Satellite remote sensing data can provide large-scale macro information, such as satellite cloud maps that reflect the distribution and movement of clouds, and surface albedo data that is closely related to photovoltaic power. Site observation data is the result of close-range monitoring, with historical power generation output data directly reflecting the past performance of renewable energy power generation. Local meteorological data such as wind speed and irradiance within the site can reflect the micro-meteorological characteristics surrounding the site. Meteorological model data (such as numerical weather forecast data) covers the changing trends of meteorological elements at the regional scale, providing meteorological background support for power forecasting.

[0015] The site attribute data includes site type identification (centralized wind power, distributed photovoltaic, agricultural photovoltaic complementarity, etc.) and information such as equipment aging degree and maintenance status. It can reflect the inherent characteristics of different types of sites and the impact of equipment operating conditions on power generation output.

[0016] After receiving the three types of key data mentioned above, they are first preprocessed, including spatiotemporal alignment and outlier processing. Spatiotemporal alignment can eliminate differences in timestamps and spatial coordinates between data from different sources, ensuring that the data is analyzed within the same spatiotemporal framework. For example, the temporal resolution of satellite remote sensing data is matched to the sampling frequency of station observation data, the grid coordinates of meteorological model data are aligned with the actual geographic extent of the target area, and station attribute data is accurately associated with the spatial location of the station to which it belongs. Outlier processing aims to eliminate abnormal data caused by equipment failure, transmission errors or extreme interference (such as negative power in station observations, noise points in satellite cloud images, and unreasonable equipment aging parameters in station attribute data), and ensure data reliability through interpolation, smoothing or elimination.

[0017] S20 extracts cloud system dynamic characteristics and surface radiation characteristics from satellite remote sensing data, extracts power time series characteristics and local meteorological fluctuation characteristics from station observation data, extracts regional meteorological trend characteristics from meteorological model data, and extracts station type characteristics and equipment status characteristics from station attribute data.

[0018] Use appropriate feature extraction networks to extract features from the preprocessed key data to transform high-dimensional, complex raw data into features with clear physical meaning or statistical laws. Specifically: Feature extraction from satellite remote sensing data: Cloud dynamic characteristics include changes in cloud morphology, movement speed, and direction. These characteristics directly affect the intensity of solar radiation reaching the Earth's surface, and thus affect photovoltaic power. Surface albedo characteristics reflect the Earth's ability to reflect solar radiation, are closely related to the irradiance received by photovoltaic modules, and are an important basis for photovoltaic power prediction. Feature extraction of station observation data: By analyzing the changing patterns of historical power generation output data (such as periodicity, trends, and mutation points), the time evolution characteristics of renewable energy power are captured, that is, the power time series characteristics are extracted; the short-term fluctuation amplitude and frequency of data such as wind speed and irradiance are extracted to reflect the immediate impact of changes in small-scale meteorological conditions around the station on power output, that is, the local meteorological fluctuation characteristics are obtained. Meteorological model data feature extraction: Regional meteorological trend features focus on the overall changing trends of meteorological elements such as temperature, air pressure, and large-scale wind speed in the target area, providing support for understanding the changing background of renewable energy power at the regional scale. For example, an increase in the overall wind speed in the region may indicate a general increase in wind power.

[0019] Extraction of site attribute data features: Site type features can reflect the inherent differences in power generation output characteristics of different types of sites (such as centralized wind power and distributed photovoltaics). For example, the sensitivity of wind power to wind speed is different from that of photovoltaics to irradiance. Equipment status features reflect the impact of non-meteorological factors such as equipment aging and maintenance status on power generation output. For example, aging equipment may lead to a decrease in power output efficiency.

[0020] It can be understood that the dynamic characteristics of cloud systems and surface radiation characteristics can be extracted through convolutional neural networks; the power timing characteristics and local meteorological fluctuation characteristics can be extracted through long-short-term memory networks; the regional meteorological trend characteristics can be extracted through attention mechanisms; and the site type characteristics and equipment status characteristics can be extracted through fully connected neural networks or attention mechanisms.

[0021] S30, using a feature fusion network to perform multi-level fusion on the extracted cloud system dynamic features, surface radiation features, power time series features, local meteorological fluctuation features, regional meteorological trend features, site type features and equipment status features to obtain fused features; wherein the number of fusion levels is determined based on at least a data feature complexity index, a prediction scenario dynamic index, and a feature correlation strength index.

[0022] This step achieves deep integration of multimodal features through feature fusion network to give full play to the complementarity of different features and enhance the representation ability of features. Specifically: First, the number of multi-level fusion layers is dynamically adjusted based on factors such as data feature complexity, the dynamic nature of the forecast scenario, and the strength of feature correlations. For example, when data features are complex (e.g., dramatic cloud changes during severe convective weather), the forecast scenario is highly dynamic (e.g., areas with multiple terrain types), or the strength of feature correlations is high and variable (e.g., cloud cover changes are highly correlated with photovoltaic power fluctuations, and this relationship changes over time), the number of fusion layers is increased to fully exploit the deep correlations between features. Conversely, the number of layers is reduced to improve fusion efficiency.

[0023] During the fusion process, through operations such as feature alignment, weight distribution and interaction, the coordination of different modal features in the temporal and spatial dimensions is achieved to avoid the "data island" phenomenon.

[0024] S40: Input the fusion features into a prediction model to obtain the predicted ultra-short-term new energy power of the target area.

[0025] The fused features obtained in step S30 integrate key information from multiple sources, including various factors affecting renewable energy power and their relationships. The fused features are then used to predict the ultra-short-term renewable energy power in the target area. Specifically, the prediction model learns the mapping relationship between the fused features and historical power data to output a forecast of renewable energy power for the next 0-4 hours.

[0026] It is understandable that the prediction model can use algorithms with strong nonlinear fitting capabilities, such as neural networks, to adapt to the complex relationship between renewable energy power and multiple factors. Taking CNN as an example, the prediction model mainly includes an input adaptation module, a local feature extraction module, a feature dimensionality reduction and enhancement module, a global feature fusion module, and a prediction output module. A brief introduction is as follows: (1) Input module: receives input fusion features (original fusion features, random features).

[0027] (2) Local feature extraction module: It captures local correlation patterns in the input fusion features (such as short-term trends in power fluctuations and local effects of meteorological factors) through convolution operations.

[0028] Convolutional layer: Use multiple convolution kernels (such as 3×3, 5×5) to perform sliding window calculations on the input fusion features. Each convolution kernel extracts a local feature (for example, the i-th convolution kernel focuses on capturing the "local correlation between cloud changes and radiation intensity").

[0029] Activation layer: Introduces nonlinearity through activation functions such as ReLU and LeakyReLU to enhance the prediction model's ability to express complex features.

[0030] (3) Feature dimensionality reduction and enhancement module: Reduce the data dimension while retaining key features, reduce the amount of calculation, and enhance the weight of important features.

[0031] Pooling layer: compresses the feature space through maximum pooling or average pooling (for example, 2×2 pooling reduces a 16×16 feature map to 8×8), while retaining significant features in local areas (such as peaks and valleys of power fluctuations).

[0032] Attention layer: Introduces a spatial attention mechanism to assign weights to the pooled feature maps (for example, assigning higher weights to the feature areas of the station equipment status) to strengthen the impact of key information on prediction.

[0033] (4) Global feature fusion module: Integrates the multi-dimensional features output by the local feature extraction module to mine global correlation patterns (such as the overall linkage between regional meteorological trends and local power fluctuations).

[0034] Fully connected layer: The high-dimensional feature map (such as (batch_size, 8, 8, 16)) is converted into a vector (batch_size, 8 × 8 × 16) by flattening, and then mapped to a low-dimensional space (such as (batch_size, 128)) through a fully connected layer to achieve global feature integration.

[0035] Residual connection: Residual blocks are added to deep CNNs to transfer shallow layer features directly to deep layers through skip connections, avoiding gradient vanishing and ensuring effective transmission of global features.

[0036] (5) Prediction output module: Map the fused global features into prediction results and output the ultra-short-term new energy power value.

[0037] Regression output layer: The predicted value is directly output through a linearly activated fully connected layer (for example, Linear(128,16) outputs the power prediction value for the next 16 time steps, assuming each step is 5 minutes, covering an ultra-short-term interval of 80 minutes).

[0038] The above prediction results can provide timely and accurate decision-making references for the power dispatching department, helping the power grid to achieve efficient absorption of new energy and stable operation.

[0039] The present invention uses a multi-level fusion algorithm to integrate multimodal data features to provide more comprehensive and targeted data support for regional ultra-short-term renewable energy power forecasts, which can significantly improve the performance and generalization ability of the prediction model, help power dispatching departments make more accurate decisions, and achieve efficient consumption of renewable energy and stable operation of the power grid.

[0040] In some embodiments, as Figure 3 As shown, the number of fusion levels is determined based on at least the data feature complexity index, the prediction scenario dynamic index, and the feature correlation strength index, including: S301, calculating the rate of change of cloud system dynamic characteristics, the fluctuation amplitude of power time series characteristics, and the variance of local meteorological fluctuation characteristics, and using them as the data characteristic complexity index; S302, calculating the terrain complexity index, meteorological disaster warning level, and new energy station distribution density of the target area, and using them as the dynamic indicators of the prediction scenario; S303, calculating the correlation coefficient between cloud system dynamic characteristics and surface radiation characteristics, the mutual information value between power time series characteristics and local meteorological fluctuation characteristics, and the matching degree between station type characteristics and power time series characteristics, and using them as the characteristic correlation strength indicators; S304: Input the data feature complexity index, the prediction scene dynamic index, and the feature association strength index into the attention weight network, output the attention weight of each index through the fully connected layer and the softmax function, and calculate the normalized score of each attention weight; S305, determining the number of levels based on the comprehensive score; wherein the comprehensive score is obtained by weighted calculation of each score.

[0041] The present invention comprehensively determines the number of fusion levels based on the data feature complexity index, the prediction scenario dynamic index, and the feature correlation strength index. The indicators are introduced as follows: (1) Data feature complexity index: This includes the rate of change of cloud system dynamic characteristics (reflecting the intensity of cloud movement and morphological changes), the amplitude of fluctuation of power time series characteristics (reflecting the stability of renewable energy power generation output), and the variance of local meteorological fluctuation characteristics (representing the discreteness of small-scale meteorological elements). This index quantifies the demand for fusion level based on the complexity of the data itself.

[0042] The rate of change of cloud dynamic characteristics: The cloud movement velocity vector is obtained by calculating the optical flow field of satellite cloud images in continuous time frames. The vector modulus is differentiated in the time dimension and the cumulative sum of the absolute values ​​is taken to obtain the rate of change of cloud morphology. Fluctuation amplitude of power time series characteristics: Calculate the range of the station's historical power output data within the sliding time window, that is, the difference between the maximum and minimum values ​​within the window; Variance of local meteorological fluctuation characteristics: Calculate the statistical variance of meteorological data such as wind speed and irradiance at the station within a preset time period.

[0043] (2) Dynamic indicators of the prediction scenario: including the terrain complexity index of the target area (reflecting the difference in the impact of regional terrain on meteorology and power generation output), the meteorological disaster warning level (reflecting the increase in prediction difficulty caused by extreme weather) and the distribution density of new energy stations (representing the complexity of collaborative modeling of stations in the region).

[0044] Terrain complexity index: Based on digital elevation model data, it calculates the weighted sum of terrain relief, slope standard deviation, and surface incision depth per unit area within the target area; Meteorological disaster warning level: Based on the warning information of rainstorms, typhoons and other disasters issued by the meteorological department, the warning level is quantified into a numerical value (e.g. blue warning is 1, yellow is 2, orange is 3, and red is 4); New energy station distribution density: Calculate the number of new energy stations per unit area in the target area.

[0045] (3) Characteristic correlation strength index: including the correlation coefficient between cloud system dynamic characteristics and surface radiation characteristics (reflecting the close correlation between meteorological elements and photovoltaic power generation output), the mutual information value between power time series characteristics and local meteorological fluctuation characteristics (reflecting the degree of influence of micrometeorology on power generation output), and the matching degree between station type characteristics and power time series characteristics (representing the differences in power generation output characteristics of different types of stations).

[0046] Correlation coefficient between cloud dynamic characteristics and surface radiation characteristics: Pearson correlation coefficient is used to calculate the correlation between cloud movement speed and surface albedo in time series; Mutual information between power time series characteristics and local meteorological fluctuation characteristics: The mutual information between power time series characteristics and meteorological fluctuation characteristics is calculated through the probability distribution after discretization. Matching degree between station type characteristics and power time series characteristics: based on the cosine similarity between the power variation patterns of different types of stations and the corresponding power time series characteristics.

[0047] Furthermore, the weight of the influence of data feature complexity, scenario dynamics, and feature correlation strength on the fusion layer varies across different scenarios. For example, in extreme weather scenarios, the impact of meteorological disaster warning levels should be significantly increased; whereas in areas with uniform terrain, the weight of the terrain complexity index can be reduced. To address this, this embodiment introduces an attention mechanism to dynamically assign weights to various indicators, making the calculation of the number of layers more tailored to actual scenario requirements and avoiding the inadequate adaptability caused by fixed weights.

[0048] Specifically, the above three types of indicators are input into the attention weight network, the nonlinear relationship between the indicators is learned through the fully connected layer, and then the normalized attention weights are output through the softmax function (the sum of each attention weight is 1), so as to realize the dynamic allocation of the importance of different indicators.

[0049] After normalizing each metric from 0 to 1, the overall score (weighted sum) is calculated based on the attention weights. Finally, the number of fusion layers (e.g., 2, 3, 4) is determined based on the score range (e.g., low, medium, high), enabling adaptive adjustment of the number of layers. For example, when the overall score S is less than 0.3, the number of fusion layers is 2; when 0.3 ≤ S < 0.7, the number of fusion layers is 3; and when S ≥ 0.7, the number of fusion layers is 4.

[0050] In some embodiments, the feature fusion network includes a feature alignment layer, a weight adaptation layer, and a cross-modal interaction layer.

[0051] In some embodiments, as Figure 4As shown in the figure, a feature fusion network is used to perform multi-level fusion of the extracted cloud system dynamic features, surface radiation features, power time series features, local meteorological fluctuation features, regional meteorological trend features, station type features, and equipment status features to obtain fused features, including: S31, the feature alignment layer maps cloud dynamic features, surface radiation features, power time series features, local meteorological fluctuation features, regional meteorological trend features, station type features, and equipment status features into a unified high-dimensional feature space. It aligns the time step with the spatial scale through spatiotemporal coordinate conversion and outputs aligned features with consistent dimensions. S32, a weight adaptation layer assigns a dynamic weight to each type of aligned features after alignment, and performs weighted concatenation of each type of aligned features based on the dynamic weight to obtain a feature matrix; S33, inputting the feature matrix into the multi-head self-attention mechanism of the cross-modal interaction layer, the multi-head self-attention mechanism mines the potential correlation between different features, performs feature fusion based on the potential correlation, and outputs the integrated fusion feature.

[0052] In some embodiments, the weight value of the dynamic weight is dynamically adjusted based on the feature association strength index, and the dynamic weights of cloud system dynamic characteristics and surface radiation characteristics, power timing characteristics and local meteorological fluctuation characteristics are strengthened through an attention mechanism.

[0053] This embodiment achieves deep integration of multimodal features through the synergy of the feature alignment layer, weight adaptation layer, and cross-modal interaction layer of the feature fusion network. Specifically: First, the feature alignment layer of the feature fusion network eliminates heterogeneity in the spatiotemporal and feature dimensions of the seven aforementioned feature types. For features of varying modalities, such as cloud dynamics (gridded satellite data), surface radiation (spatially distributed data), and power temporal characteristics (time series data), the coordinate conversion module first unifies the spatial coordinates to the target region's geographic coordinate system (e.g., WGS84). A temporal interpolation module then resamples features with varying sampling frequencies to the same time step (e.g., 10 minutes per step) to ensure consistency in spatiotemporal scales. An embedding layer then maps these features to a unified high-dimensional feature space (e.g., a 256-dimensional vector), outputting aligned features with matching dimensions. This process avoids information fragmentation caused by spatiotemporal misalignment or dimensionality differences.

[0054] Next, the weight adaptation layer of the feature fusion network dynamically adjusts the contribution of each feature type based on the feature correlation strength index derived in the previous steps. First, using metrics such as the correlation coefficient and mutual information value calculated previously, an initial weight is assigned to each aligned feature type. Second, for pairs of cloud dynamic features and surface radiation characteristics (strongly correlated feature pairs), and power time series features and local meteorological fluctuation characteristics (strongly influential feature pairs), an attention mechanism is used to calculate correlation scores between these pairs. These scores are normalized and used as reinforcement weights, which are then added to the initial weights. Finally, a softmax function is used to output the normalized dynamic weights, ensuring that key features (such as cloud characteristics that affect photovoltaic power) receive higher weight in the fusion process. This adaptive weight allocation based on feature contribution avoids interference from irrelevant features.

[0055] Based on the dynamic weights obtained above, the seven types of alignment features are weighted element by element and then spliced ​​into a feature matrix according to the channel dimension.

[0056] Next, the number of heads of the multi-head self-attention mechanism is determined based on the number of fusion layers determined above. Specifically, the number of heads of the multi-head self-attention mechanism is positively correlated with the number of fusion layers. For example, when the number of fusion layers is 2, the number of heads of the multi-head self-attention mechanism is set to 2, focusing on spatiotemporal correlation and attribute-power output correlation respectively; when the number of fusion layers is 3, the number of heads of the multi-head self-attention mechanism is increased to 4, and the new heads are used to capture the cross-domain correlation between meteorological pattern data and station attribute data; when the number of fusion layers is 4, the number of heads of the multi-head self-attention mechanism is further expanded to 8, and the residual attention mechanism between layers is introduced to enable high-level heads to integrate the basic correlation features extracted by low-level heads.

[0057] The output dimension of each attention head is dynamically adjusted based on the number of fusion levels, calculated as: head dimension = total feature dimension / (number of heads × number of fusion levels). This setting ensures that as the fusion level increases, the feature fusion network can discover more complex feature association patterns while maintaining a balance between computational efficiency and representational capabilities.

[0058] The cross-modal interaction layer exploits deep connections between features through a multi-head self-attention mechanism. Specifically, the weighted seven feature categories are concatenated into a feature matrix and fed into a multi-head self-attention module. Each attention head focuses on learning correlation patterns along different dimensions. For example, the first attention head captures the spatial transmission relationship between regional meteorological trends and local meteorological fluctuations. The second attention head learns the attribute-to-power generation output mapping between site type, equipment status, and power time series. The third attention head analyzes the temporal linkage between cloud dynamics and surface radiation. The concatenation layer then integrates the outputs of these multiple heads. After residual connections and layer normalization, the output is a fused feature that incorporates cross-modal information. This process makes implicit connections between features (such as the meteorological-power generation output relationship influenced by terrain) explicit. The cross-modal interaction layer achieves deep feature interaction through multiple layers of Transformer blocks, each containing self-attention and feed-forward neural networks. Furthermore, LSTM / GRU layers can be added before the cross-modal interaction layer to handle temporal dependencies.

[0059] This embodiment uses the above-mentioned progressive processing of alignment, weighting, and interaction to not only solve the heterogeneity problem of multi-source features, but also strengthen key information through dynamic weights and deep correlation mining. The final output fusion features can comprehensively reflect the various factors affecting the power of new energy and their correlations, providing high-quality input for subsequent predictions.

[0060] In some embodiments, as Figure 5 As shown, the fusion features are input into the prediction model to obtain the predicted ultra-short-term new energy power of the target area, including: S41, calculating an equipment status score based on the station type characteristics, determining a matching volatility coefficient based on the equipment status characteristics, calculating a station attribute evaluation value based on the equipment status score and the volatility coefficient, and matching the station attribute evaluation value to obtain an intensity interval of the random feature; S42, generating random features based on the intensity interval, integrating the random features into the fusion features and inputting them into a prediction model together to obtain the predicted ultra-short-term new energy power of the target area.

[0061] The actual power generation output of renewable energy stations is not only affected by external factors such as weather and topography, but also by random fluctuations due to the characteristics of the equipment itself. For example, aging photovoltaic panels can cause instantaneous power jumps, and wind turbine gearbox wear can cause unstable power output. These inherent irregular disturbances are difficult to fully model using multimodal features. Directly inputting fused features into the prediction model can easily ignore these random fluctuations, resulting in overly smoothed predictions that fail to reflect unexpected deviations in actual operation.

[0062] To address the above technical issues, this embodiment introduces random features based on site attribute evaluation values ​​to simulate power generation fluctuations caused by inherent equipment instability. For example, the higher the site attribute evaluation value (the worse the equipment condition, the greater the volatility), the stronger the random feature. This allows the prediction model to improve its adaptability to irregular fluctuations based on the mapping relationship between multimodal features and random disturbances and actual power generation output, ultimately reducing the deviation between the predicted value and the actual power generation output.

[0063] The random features are added to the original fusion features element by element to form enhanced fusion features. This not only retains the core correlation of multimodal information, but also injects random disturbances that match the site attributes. It can simulate the power generation output fluctuations caused by the instability of the equipment itself in actual operation, and improve the accuracy of the predicted ultra-short-term new energy power in the target area.

[0064] It is understandable that the prediction model is pre-trained based on the fusion features that are fused with random features. The specific training process is roughly as follows: Step 1: Select the fusion features of the same period in history, generate corresponding random features according to the determined random feature intensity interval, and fuse them with the historical fusion features to form a pre-training sample set containing random perturbations of different intensities; Step 2: Input the pre-trained sample set into the initial prediction model and iterate the training with the Adam optimizer using the mean squared error loss function until the model loss on the validation set converges. In each round of training, random features of varying strengths are randomly selected and injected into the fusion features to enable the model to learn the mapping rules under different device fluctuation scenarios. In step 3, the pre-trained model parameters are used as initial values ​​and fine-tuned using the fused features (including corresponding random features) collected in real time. The model is optimized by reducing the learning rate (for example, to 1 / 10 of the pre-training stage) to adapt it to the latest equipment status and meteorological characteristics of the current area.

[0065] In some embodiments, the calculating of the equipment status score based on the station type characteristics and the determining of the matching volatility coefficient based on the equipment status characteristics include: The weight of each scoring dimension is determined based on the site type characteristics, and the equipment status score is calculated by combining various quantitative indicators; the volatility coefficient is determined based on the equipment status characteristics and a preset matching relationship.

[0066] Different types of new energy stations have significant differences in the core influencing factors of their equipment. Therefore, the weights of each scoring dimension need to be dynamically allocated according to the characteristics of the station type (such as centralized wind power, distributed photovoltaics, agricultural and photovoltaic complementarity, etc.).

[0067] For example: The equipment status of centralized wind farms is more dependent on the "degree of wind turbine blade wear" and "gearbox operating temperature", so the weights of these two quantitative indicators (such as 0.3 and 0.25) are higher than the "degree of cable aging" (weight 0.15); distributed photovoltaic stations are more affected by "component cleanliness" and "inverter conversion efficiency", and the corresponding indicator weights (such as 0.3 and 0.28) are significantly higher than other dimensions.

[0068] After determining the weight of each scoring dimension, the equipment status score (ranging from 0 to 1, with lower scores indicating more stable equipment status) is calculated by weighted summation of various quantitative indicators (such as equipment operating years, fault repair time, performance degradation rate, etc.).

[0069] Equipment status characteristics (such as equipment age, historical failure frequency, and most recent maintenance record) directly reflect its operational stability. The volatility coefficient (ranging from 0.1 to 0.8, with a higher coefficient indicating a greater potential for fluctuation in the equipment's own power generation output) must be determined based on a preset matching relationship.

[0070] For example, in the preset rules, the volatility coefficient for photovoltaic modules with a service life of less than 5 years and no major failures is 0.2; the volatility coefficient for modules with a service life of 10 years or more and an average of 3 failures per year is 0.6; for wind turbine equipment, the coefficient for gearbox oil testing that passes and no abnormal vibration in the last three months is 0.3; the coefficient for bearing wear that exceeds the standard and has frequent historical fluctuation records is 0.7.

[0071] Through the above matching relationship, the qualitative description of the equipment status characteristics is converted into a quantitative coefficient to accurately describe the degree of influence of the equipment's own characteristics on the fluctuation of power generation output.

[0072] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A regional ultra-short-term new energy power prediction method based on multimodal information fusion, characterized in that: The steps include: Obtain satellite remote sensing data, station observation data, meteorological model data, and station attribute data of the target area, and perform spatiotemporal alignment and outlier processing; Extract cloud system dynamic characteristics and surface radiation characteristics from satellite remote sensing data, extract power time series characteristics and local meteorological fluctuation characteristics from station observation data, extract regional meteorological trend characteristics from meteorological model data, and extract station type characteristics and equipment status characteristics from station attribute data. A feature fusion network is used to perform multi-level fusion of the extracted cloud system dynamic features, surface radiation features, power time series features, local meteorological fluctuation features, regional meteorological trend features, station type features, and equipment status features to obtain fused features. The number of fusion levels is determined at least based on the data feature complexity index, the prediction scenario dynamic index, and the feature correlation strength index. The fusion features are input into a prediction model to obtain the predicted ultra-short-term new energy power of the target area.

2. The method for regional ultra-short-term renewable energy power prediction based on multimodal information fusion according to claim 1, characterized in that: The number of fusion levels is determined based on at least a data feature complexity index, a prediction scenario dynamics index, and a feature correlation strength index, including: Calculate the rate of change of cloud system dynamic characteristics, the fluctuation amplitude of power time series characteristics, and the variance of local meteorological fluctuation characteristics, and use them as the complexity index of the data characteristics; Calculate the terrain complexity index, meteorological disaster warning level, and new energy station distribution density of the target area and use them as dynamic indicators of the prediction scenario; Calculate the correlation coefficient between cloud system dynamic characteristics and surface radiation characteristics, the mutual information value between power time series characteristics and local meteorological fluctuation characteristics, and the matching degree between station type characteristics and power time series characteristics, and use them as the characteristic correlation strength indicators; Input the data feature complexity index, the prediction scene dynamic index, and the feature association strength index into the attention weight network, output the attention weight of each index through the fully connected layer and the softmax function, and calculate the normalized score of each attention weight; The number of levels is determined based on a comprehensive score, wherein the comprehensive score is obtained by weighted calculation of each normalized score.

3. The method for regional ultra-short-term new energy power prediction based on multimodal information fusion according to claim 2, characterized in that: The feature fusion network includes a feature alignment layer, a weight adaptation layer, and a cross-modal interaction layer.

4. The method for regional ultra-short-term new energy power prediction based on multimodal information fusion according to claim 3 is characterized by: A feature fusion network is used to perform multi-level fusion of the extracted cloud system dynamic features, surface radiation features, power time series features, local meteorological fluctuation features, regional meteorological trend features, station type features, and equipment status features to obtain fused features, including: The feature alignment layer maps cloud dynamic features, surface radiation features, power time series features, local meteorological fluctuation features, regional meteorological trend features, station type features, and equipment status features into a unified high-dimensional feature space. It aligns the time step with the spatial scale through spatiotemporal coordinate conversion and outputs aligned features with consistent dimensions. The weight adaptation layer assigns a dynamic weight to each type of aligned features after alignment, and performs weighted concatenation of each type of aligned features based on the dynamic weight to obtain a feature matrix; The feature matrix is ​​input into the multi-head self-attention mechanism of the cross-modal interaction layer. The multi-head self-attention mechanism mines the potential correlation between different features, performs feature fusion based on the potential correlation, and outputs the integrated fusion feature.

5. The method for regional ultra-short-term new energy power prediction based on multimodal information fusion according to claim 4 is characterized by: The weight value of the dynamic weight is dynamically adjusted based on the feature association strength index, and the dynamic weights of cloud system dynamic characteristics and surface radiation characteristics, power timing characteristics and local meteorological fluctuation characteristics are strengthened through the attention mechanism.

6. The method for regional ultra-short-term new energy power prediction based on multimodal information fusion according to claim 2, characterized in that: The fusion features are input into the prediction model to obtain the predicted ultra-short-term new energy power of the target area, including: Calculating an equipment status score based on the station type characteristics, determining a matching volatility coefficient based on the equipment status characteristics, calculating a station attribute evaluation value based on the equipment status score and the volatility coefficient, and matching the station attribute evaluation value to obtain an intensity interval of the random feature; A random feature is generated based on the intensity interval, and the random feature is integrated into the fusion feature and then input into the prediction model together to obtain the predicted ultra-short-term new energy power of the target area.

7. The method for regional ultra-short-term new energy power prediction based on multimodal information fusion according to claim 6, characterized in that: The equipment status score is calculated based on the site type characteristics, and the matching volatility coefficient is determined based on the equipment status characteristics, including: The weight of each scoring dimension is determined based on the site type characteristics, and the equipment status score is calculated by combining various quantitative indicators; the volatility coefficient is determined based on the equipment status characteristics and a preset matching relationship.

Images