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

Abstract

The application belongs to the technical field of power prediction. A regional ultra-short-term new energy power prediction method based on multi-modal information fusion is provided, which comprises: acquiring satellite remote sensing data, station observation data, meteorological model data and station attribute data of a target region, and performing time-space alignment and outlier processing; feature extraction is performed from the above data; a multi-level fusion network is used to perform multi-level fusion on the extracted features to obtain fused features; the fused features are input into a prediction model to obtain the predicted ultra-short-term new energy power of the target region. The application integrates multi-modal data features by using a multi-level fusion algorithm, provides more comprehensive and targeted data support for regional ultra-short-term new energy power prediction, can significantly improve the performance and generalization ability of the prediction model, and helps the power dispatching department to make more accurate decisions, realizes efficient consumption of new energy and stable operation of the power grid.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power prediction, in particular to a regional ultra-short-term new energy power prediction method based on multi-modal information fusion. BACKGROUND

[0002] Under the background of global energy transformation, new energy power generation, as a core component of clean energy, continues to rise in the penetration rate of power systems. For provincial regions, the large-scale centralized grid connection of new energy brings severe challenges to the safe and stable operation of power systems, and the ultra-short-term new energy power generation output prediction (0-4 hours in the future) as a key support for power dispatching decision-making, its accuracy is directly related to the power grid frequency modulation, power market transactions and new energy consumption efficiency.

[0003] The provincial region is vast in territory, with complex and diverse topography, covering plains, mountains, basins, and coastal areas, and the meteorological conditions show significant spatial heterogeneity. New energy power generation (especially photovoltaic and wind power) is highly dependent on meteorological elements such as light intensity, wind speed, temperature, and cloud cover, and the spatial and temporal distribution of these elements varies greatly in different regions within the province. For example, the terrain sheltering in mountainous areas can cause sudden changes in local wind speed, the sea-land wind circulation in coastal areas can cause fluctuations in wind power generation output, and the rapid migration of clouds can cause sharp 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 of new energy power generation output throughout the province.

[0004] Currently, provincial ultra-short-term new energy power generation output prediction mainly relies on three types of data: first, station observation data, including historical power generation of new energy stations, wind speed, and irradiance collected by station meteorological stations; second, meteorological model data, i.e. grid meteorological elements output by numerical weather prediction (NWP); third, satellite remote sensing data, such as satellite cloud images, MODIS land albedo, etc. However, the existing technology has obvious limitations in data utilization and fusion, mainly in:

[0005] From the data source, although the station observation data can reflect the fine information of a single point, it is difficult to cover all new energy-rich regions in the province due to the spatial distribution density of observation stations, and is easily affected by equipment failure, transmission delay, etc., lacking data integrity and timeliness. Although meteorological model data has the ability to cover the whole province, the spatial resolution is low (usually 3-10 kilometers), making it difficult to depict local meteorological anomalies caused by micro-topography such as valleys and building groups, and the prediction error will accumulate as the prediction period extends. Satellite remote sensing data can provide macro information such as cloud movement and surface radiation over a large area, but it has low time resolution (such as updating every 15-30 minutes) and is affected by cloud and rain shielding, making it difficult to meet the demand for high-frequency dynamic monitoring of ultra-short-term prediction.

[0006] From the perspective of fusion methods, traditional prediction models mostly use simple splicing or weighted superposition to process multi-source data, failing to deeply mine the internal correlation of different modal data. For example, machine learning-based models often train separately using historical data from the station, ignoring the future irradiance trend implied in satellite cloud images; although physical models combine meteorological model data, they do not effectively use station observation data to correct prediction bias in real time. This "data island" phenomenon prevents the complementary nature of multi-source information from being fully utilized, and in complex weather conditions (such as severe convective weather and rapid temperature change processes), prediction errors often greatly exceed the acceptable range of actual scheduling.

[0007] In addition, provincial new energy stations have various types (centralized wind power, distributed photovoltaic, and agricultural-photovoltaic complementary, etc.), and the power generation characteristics of different types of stations differ significantly, and are affected by non-weather factors such as equipment aging and maintenance status. Existing prediction methods mostly focus on a single energy type or homogeneous areas, lack the ability to model multiple types of new energy stations in the province, and are difficult to meet the prediction needs of the provincial power grid for the total power generation of new energy stations.

[0008] Therefore, how to break through the limitations of single-source data, deeply integrate satellite remote sensing, station observation, and meteorological model data through multi-modal fusion algorithms, fully exploit the complementary value of different data in the time and space scales and information dimensions, and improve the accuracy and robustness of provincial ultra-short-term new energy power generation prediction is a key problem that needs to be solved in the field of new energy power systems. SUMMARY

[0009] To solve the technical problems in the background art, the present application provides a regional ultra-short-term new energy power prediction method based on multi-modal information fusion.

[0010] The present application provides a regional ultra-short-term new energy power prediction method based on multi-modal information fusion, comprising the following steps:

[0011] Obtain satellite remote sensing data, station observation data, meteorological model data, and station attribute data of the target region, and perform time and space alignment and outlier processing;

[0012] Extract cloud system dynamic features and surface radiation features from satellite remote sensing data, power time series features and local meteorological fluctuation features from station observation data, regional meteorological trend features from meteorological model data, and station type features and equipment status features from station attribute data;

[0013] The extracted cloud system dynamic features, ground radiation features, power time sequence features, local meteorological fluctuation features, regional meteorological trend features, station type features and equipment state features are fused by using a feature fusion network to obtain fused features, wherein the number of fusion layers is determined based on at least a data feature complexity index, a prediction scene dynamic index and a feature correlation strength index;

[0014] The fused features are input into a prediction model to obtain predicted super-short-term new energy power of the target region.

[0015] The present application integrates multi-modal data features by using a multi-level fusion algorithm, provides more comprehensive and targeted data support for regional super-short-term new energy power prediction, can significantly improve the performance and generalization ability of the prediction model, and helps the power dispatching department to make more accurate decisions, realizes efficient consumption of new energy and stable operation of the power grid. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of a regional super-short-term new energy power prediction method based on multi-modal information fusion disclosed by the embodiments of the present application;

[0017] Figure 2 is a scene schematic diagram of the present application scheme;

[0018] Figure 3 is a schematic diagram of the determination process of the number of fusion layers disclosed by the embodiments of the present application;

[0019] Figure 4 is a flowchart of obtaining fused features by multi-level fusion disclosed by the embodiments of the present application;

[0020] Figure 5 is a flowchart of predicting the super-short-term new energy power of the target region based on the fused features disclosed by the embodiments of the present application. DETAILED DESCRIPTION

[0021] Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, which include various details of the embodiments of the present disclosure to help understanding, and should be considered as merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Also, in order to be clear and concise, the description below omits the description of well-known functions and structures.

[0022] As shown in Figure 1 , Figure 2 , the embodiments of the present application disclose a regional super-short-term new energy power prediction method based on multi-modal information fusion, comprising the following steps:

[0023] S10, obtain satellite remote sensing data, station observation data, meteorological model data, and station attribute data of the target region, and perform spatio-temporal alignment and outlier processing.

[0024] For a target region (e.g., a provincial division), the above three types of key data are synchronously obtained. Among them, satellite remote sensing data can provide macro information on a large scale, such as satellite cloud images that can reflect the distribution and movement of cloud layers, and surface albedo data that is closely related to photovoltaic power. Station observation data is a close-range monitoring result, and historical power generation data directly reflects the past performance of new energy power generation. Local meteorological data such as station wind speed and irradiance can reflect the micro-meteorological characteristics around the station. Meteorological model data (such as numerical weather prediction data) covers the regional scale of meteorological element changes, providing meteorological background support for power prediction.

[0025] Station attribute data includes station type identification (centralized wind power, distributed photovoltaic, and agricultural-photovoltaic complementary, etc.), and information such as equipment aging degree and maintenance status, which can reflect the inherent characteristics of different types of stations and the influence of equipment operating conditions on power generation.

[0026] After receiving the above three types of key data, pre-processing operations are performed, including spatio-temporal alignment and outlier processing. Among them, spatio-temporal alignment can eliminate the differences in time stamp and spatial coordinates of data from different sources, ensuring that the data is analyzed in the same spatio-temporal framework. For example, match the time resolution of satellite remote sensing data with the sampling frequency of station observation data, correspond the grid coordinates of meteorological model data with the actual geographical range of the target region, and accurately associate the station attribute data with the spatial position of the station.

[0027] Outlier processing aims to eliminate abnormal data caused by equipment failure, transmission error or extreme interference (such as negative power in station observation, noise points in satellite cloud images, and unreasonable equipment aging parameters in station attribute data), ensuring the reliability of the data through interpolation, smoothing or elimination.

[0028] S20, extract cloud system dynamic features and surface radiation features from satellite remote sensing data, power time series features and local meteorological fluctuation features from station observation data, regional meteorological trend features from meteorological model data, and station type features and equipment state features from station attribute data.

[0029] Use a suitable feature extraction network to extract features from the pre-processed key data above, to convert high-dimensional, complex raw data into features with clear physical meaning or statistical rules. Specifically:

[0030] Satellite remote sensing data feature extraction: Cloud system dynamic features include morphological changes of cloud layers, moving speed and direction, etc. These features directly affect the intensity of solar radiation reaching the ground, and in turn affect photovoltaic power; The surface albedo feature reflects the ability of the ground to reflect solar radiation, which is closely related to the irradiance received by the photovoltaic module, and is an important basis for photovoltaic power prediction.

[0031] Field station observation data feature extraction: By analyzing the change rules of historical power generation data (such as periodicity, trend, and mutation points), the temporal evolution characteristics of new energy power are captured, that is, the power time series features are obtained; The short-term fluctuation amplitude and frequency of wind speed, irradiance and other data are extracted, which reflects the immediate impact of the changes of small-scale meteorological conditions around the field station on power output, that is, the local meteorological fluctuation features are obtained.

[0032] Meteorological model data feature extraction: Regional meteorological trend features focus on the overall change trend of temperature, pressure, and large-scale wind speed in the target region, providing support for understanding the change background of new energy power at the regional scale, for example, the overall wind speed enhancement in the region may indicate the general rise of wind power.

[0033] Field station attribute data feature extraction: Field station type features can reflect the inherent differences in power generation output characteristics of different types of field stations (such as centralized wind power and distributed photovoltaic power), for example, the sensitivity of wind power to wind speed is different from that of photovoltaic power to irradiance; Equipment status features reflect the influence of non-meteorological factors such as equipment aging degree and maintenance situation on power generation output, for example, aging equipment may lead to a decrease in power output efficiency.

[0034] It can be understood that cloud system dynamic features and surface radiation features can be extracted by convolutional neural networks; and power time series features and local meteorological fluctuation features can be extracted by long short-term memory networks; and regional meteorological trend features can be extracted by attention mechanisms; and field station type features and equipment status features can be extracted by fully connected neural networks or attention mechanisms, etc.

[0035] 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, field station type features, and equipment status features, to obtain fusion features; wherein the number of fusion layers is determined based on at least data feature complexity indicators, prediction scene dynamic indicators, and feature correlation strength indicators.

[0036] This step realizes the deep integration of multi-modal features through a feature fusion network, so as to fully exert the complementarity of different features and improve the representation ability of the features. Specifically:

[0037] First, the number of fusion layers is dynamically adjusted according to factors such as data feature complexity, prediction scene dynamics, and feature correlation strength. For example, when the data feature is complex (such as severe convective weather with rapid cloud system changes), the prediction scene is highly dynamic (such as areas with multiple terrains), and the feature correlation is strong and variable (such as a high correlation between cloud changes and photovoltaic power fluctuations, and the relationship changes over time), the number of fusion layers is increased to fully exploit the deep associations between features; otherwise, the number of layers is reduced to improve fusion efficiency.

[0038] During the fusion process, feature alignment, weight distribution, and interaction operations are used to realize the cooperation of different modal features in the time and space dimensions, avoiding the "data island" phenomenon.

[0039] S40, input the fusion features into a prediction model to obtain a predicted ultra-short-term new energy power of the target area.

[0040] The fusion features obtained through step S30 have integrated the key information of multi-source data and contain various factors affecting new energy power and their associated relationships. At this time, the prediction of the ultra-short-term new energy power of the target area is completed using the fused fusion features, that is, the prediction model learns the mapping relationship between the fusion features and the historical power data, and outputs the new energy power prediction value for the next 0-4 hours.

[0041] It can be understood that the prediction model can use algorithms such as neural networks that have strong non-linear fitting capabilities to adapt to the complex relationship between new 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 dimension reduction and enhancement module, a global feature fusion module, and a prediction output module, which are briefly introduced as follows:

[0042] (1) Input module: receives the input fusion features (original fusion features, random features).

[0043] (2) Local feature extraction module: captures local correlation patterns in the input fusion features (such as short-term trends of power fluctuations, local effects of meteorological factors) through convolution operations.

[0044] Convolution layer: uses multiple convolution kernels (such as 3x3, 5x5) to perform sliding window calculation on the input fusion features, and each convolution kernel extracts a local feature (for example, the i-th convolution kernel focuses on capturing the local correlation between cloud system changes and radiation intensity).

[0045] Activation layer: introduces nonlinearity through ReLU, LeakyReLU, etc. activation function to enhance the prediction model's ability to express complex features.

[0046] (3) Feature dimension reduction and strengthening module: reduce the data dimension under the premise of retaining key features, reduce the amount of calculation, and strengthen the weight of important features.

[0047] Pooling layer: compress the feature space through Max Pooling or Average Pooling (such as 2x2 pooling reduces 16x16 feature map to 8x8), while retaining the significant features of local areas (such as the peak / valley of power fluctuations).

[0048] Attention layer: introduce a spatial attention mechanism to assign weights to the pooled feature map (for example, give higher weight to the station equipment state feature area), to strengthen the impact of key information on prediction.

[0049] (4) Global feature fusion module: integrate multi-dimensional features output by the local feature extraction module, and mine global correlation patterns (such as the overall linkage of regional meteorological trends and local power fluctuations).

[0050] Fully connected layer: convert high-dimensional feature map (such as (batch_size,8,8,16)) to vector (batch_size,8x8x16) through flattening, and then map it to low-dimensional space (such as (batch_size,128)) through fully connected layer, to realize the integration of global features.

[0051] Residual connection: add a residual block (Residual Block) to the deep CNN, which directly transmits shallow features to deep layers through a skip connection, avoiding gradient vanishing and ensuring effective transmission of global features.

[0052] (5) Prediction output module: map the fused global features to the prediction result, and output the value of the ultra-short-term new energy power.

[0053] Regression output layer: directly output the prediction value through a linearly activated fully connected layer (such as Linear(128,16), output the power prediction value for the next 16 time steps, assuming 5 minutes per step, covering an 80-minute ultra-short-term interval).

[0054] The above prediction results can provide timely and accurate decision-making reference for power dispatching departments, helping the power grid to achieve efficient consumption and stable operation of new energy.

[0055] The present application uses a multi-level fusion algorithm to integrate multi-modal data features, providing more comprehensive and targeted data support for regional ultra-short-term new energy power prediction, which can significantly improve the performance and generalization ability of the prediction model, helping power dispatching departments to make more accurate decisions, achieving efficient consumption of new energy and stable operation of the power grid.

[0056] In some embodiments, as shown in Figure 3 The fusion layer number is determined based on at least a data feature complexity index, a prediction scene dynamic index, and a feature correlation strength index, including:

[0057] S301, a change rate of cloud system dynamic characteristics, a fluctuation amplitude of power time sequence characteristics, and a variance of local weather fluctuation characteristics are calculated and taken as the data feature complexity index;

[0058] S302, a terrain complexity index of the target region, a weather disaster warning level, and a new energy station distribution density are calculated and taken as the prediction scene dynamic index;

[0059] S303, a correlation coefficient of cloud system dynamic characteristics and ground surface radiation characteristics, a mutual information value of power time sequence characteristics and local weather fluctuation characteristics, and a matching degree of station type characteristics and power time sequence characteristics are calculated and taken as the feature correlation strength index;

[0060] S304, the data feature complexity index, the prediction scene dynamic index, and the feature correlation strength index are input into an attention weight network, and an attention weight of each index is output through a fully connected layer and a softmax function, and a normalized score of each attention weight is calculated;

[0061] S305, the layer number is determined based on a comprehensive score, wherein the comprehensive score is calculated by weighting each score.

[0062] The fusion layer number is determined based on a data feature complexity index, a prediction scene dynamic index, and a feature correlation strength index, and each index is introduced as follows:

[0063] (1) The data feature complexity index includes a change rate of cloud system dynamic characteristics (reflecting the intensity of cloud movement and morphological change), a fluctuation amplitude of power time sequence characteristics (reflecting the stability of new energy power generation output), and a variance of local weather fluctuation characteristics (reflecting the dispersion of small-scale meteorological elements). This index quantifies the demand for fusion layer from the complexity of data itself.

[0064] The change rate of cloud system dynamic characteristics: the optical flow field of the satellite cloud image in the continuous time frame is calculated to obtain the cloud movement speed vector, the derivative of the vector module length in the time dimension is taken and the absolute value of the cumulative sum is obtained, and the cloud morphological change rate is obtained.

[0065] The fluctuation amplitude of power time sequence characteristics: the range of the historical power generation output data of the station in the sliding time window is calculated, that is, the difference between the maximum value and the minimum value in the window.

[0066] The variance of local weather fluctuation characteristics: the statistical variance of meteorological data such as wind speed and irradiance of the station in the preset time period is calculated.

[0067] (2) Predicted scene dynamic indicators: including terrain complexity index of target area (reflecting the difference in the influence of regional terrain on meteorology and power generation output), meteorological disaster warning level (reflecting the increase in prediction difficulty due to extreme weather), and new energy station distribution density (characterizing the complexity of regional station collaborative modeling).

[0068] Terrain complexity index: based on digital elevation model data, calculate the terrain undulation, slope standard deviation and weighted sum of surface cutting depth per unit area in the target area;

[0069] Meteorological disaster warning level: according to the disaster warning information of rainstorm, typhoon and other disasters released by the meteorological department, the warning level is quantified as a numerical value (such as blue warning for 1, yellow for 2, orange for 3, and red for 4);

[0070] New energy station distribution density: calculate the number of new energy stations per unit area in the target area.

[0071] (3) Feature correlation strength indicators: including the correlation coefficient of cloud system dynamic features and surface radiation features (reflecting the close relationship between meteorological elements and photovoltaic power generation output), the mutual information value of power time series features and local meteorological fluctuation features (reflecting the influence of microclimate on power generation output), and the matching degree of station type features and power time series features (characterizing the difference in power generation output characteristics of different types of stations).

[0072] Correlation coefficient of cloud system dynamic features and surface radiation features: using Pearson correlation coefficient, calculate the correlation between cloud system moving speed and surface albedo in time series;

[0073] Mutual information value of power time series features and local meteorological fluctuation features: through the probability distribution after discretization, calculate the mutual information between power time series features and meteorological fluctuation features;

[0074] Matching degree of station type features and power time series features: based on the cosine similarity of power change rule of different types of stations and corresponding power time series features.

[0075] At the same time, under different scenarios, the influence weight of data feature complexity, scene dynamics and feature correlation strength on fusion level is different. For example, in the extreme weather scenario, the influence of meteorological disaster warning level should be significantly improved; while in the single region, the weight of terrain complexity index can be reduced. In view of this situation, the attention mechanism is introduced to dynamically allocate the weight of each indicator, so that the layer number calculation is more in line with the actual scene demand, and the problem of insufficient adaptability caused by fixed weight is avoided.

[0076] Specifically, the three types of indicators mentioned above are input into the attention weight network. The non-linear 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.

[0077] After normalizing each indicator to 0-1, a comprehensive score (weighted summation) is calculated based on attention weights. Finally, the number of fusion layers (e.g., 2, 3, 4 layers) is determined according to the score range (e.g., low, medium, high), achieving adaptive adjustment of the number of layers. For example, when the comprehensive score S < 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.

[0078] In some embodiments, the feature fusion network includes a feature alignment layer, a weight adaptation layer, and a cross-modal interaction layer.

[0079] In some embodiments, such as Figure 4 As shown, a feature fusion network is used to perform multi-level fusion of extracted 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 to obtain fused features, including:

[0080] 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 to a unified high-dimensional feature space. It achieves alignment of time step and spatial scale through spatiotemporal coordinate transformation and outputs aligned features with consistent dimensions.

[0081] S32, the weight adaptive layer assigns dynamic weights to each type of alignment feature after alignment, and performs weighted concatenation of each type of alignment feature based on the dynamic weights to obtain a feature matrix;

[0082] S33, 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 fused features.

[0083] In some embodiments, the weight values ​​of the dynamic weights are dynamically adjusted based on the feature association strength index, and the dynamic weights of cloud dynamic features and surface radiation features, power time series features and local meteorological fluctuation features are strengthened through an attention mechanism.

[0084] This embodiment achieves deep integration of multimodal features through the synergistic effect of the feature alignment layer, weight adaptation layer, and cross-modal interaction layer of the feature fusion network. Specifically:

[0085] Firstly, the feature alignment layer of the feature fusion network eliminates the heterogeneity of the above seven types of features in the space-time dimension and the feature dimension. For different forms of features such as cloud dynamic features (raster satellite data), surface radiation features (spatial distributed data), and power time series features (time series data), the spatial coordinates are first unified to the geographic coordinate system (such as WGS84 coordinate system) through the coordinate conversion module, and the features with different sampling frequencies are resampled to the same time step (such as 10 minutes / step) through the time interpolation module to ensure the consistency of the space-time scale; then the embedding layer is used to map the features of different types to a unified high-dimensional feature space (such as 256-dimensional vector), and the output dimension of the aligned features is matched. In this way, the information fragmentation caused by space-time misalignment or dimension difference can be avoided.

[0086] Then, the weight adaptive layer of the feature fusion network dynamically adjusts the contribution of each type of feature based on the feature correlation strength index obtained in the previous step. Firstly, the correlation coefficient, mutual information value and other indicators calculated in the previous step are used to assign an initial weight to each type of aligned feature; secondly, for the cloud dynamic feature and the surface radiation feature (strongly correlated feature pair), and the power time series feature and the local meteorological fluctuation feature (strongly influenced feature pair), the attention mechanism is used to calculate the correlation score between the feature pairs, and the normalized score is used as the reinforcement weight and added to the initial weight; finally, the softmax function is used to output the normalized dynamic weight, ensuring that the key features (such as cloud features affecting photovoltaic power) occupy a higher weight in the fusion. The adaptive weight allocation based on feature contribution in this step can avoid irrelevant feature interference.

[0087] Based on the dynamic weight obtained above, the seven types of aligned features are weighted element by element and concatenated into a feature matrix in the channel dimension.

[0088] Then, the number of heads of the multi-head self-attention mechanism is determined based on the number of fusion layers determined in the previous step. 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, which focuses on the space-time correlation and the attribute-power output correlation; when the number of fusion layers is 3, the number of heads of the multi-head self-attention mechanism increases to 4, and a new head is added to capture the cross-domain correlation between the meteorological pattern data and the site 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 a residual attention mechanism between layers is introduced, which enables the high-level heads to integrate the basic correlation features extracted by the low-level heads.

[0089] The output dimension of each attention head is dynamically adjusted according to the number of fusion layers, and the calculation formula is: head dimension = total feature dimension / (number of multi-heads x number of fusion layers). Through this setting, the feature fusion network can extract more complex feature correlation patterns as the number of fusion layers increases, while maintaining the balance between computational efficiency and representation ability.

[0090] The cross-modal interaction layer mines deep correlations between features through a multi-head self-attention mechanism. Specifically, the weighted seven types of features are concatenated into a feature matrix, which is input into a multi-head self-attention module. Each attention head focuses on learning different dimensional correlation patterns: for example, the first attention head captures the spatial conduction relationship between regional meteorological trend features and local meteorological fluctuation features, the second attention head learns the attribute-power output mapping relationship between station type features, equipment state features, and power time series features, and the third attention head analyzes the time series linkage relationship between cloud system dynamic features and surface radiation features; then the multi-head outputs are integrated through a concatenation layer, and after residual connection and layer normalization processing, the fusion features that integrate cross-modal information are output. In this way, the implicit correlations between features (such as the meteorological-power generation relationship under the influence of terrain) can be made explicit. It can be understood that the cross-modal interaction layer can realize feature deep interaction through multiple Transformer blocks, each of which includes self-attention and feedforward neural networks; and an LSTM / GRU can also be added before the cross-modal interaction layer to handle temporal dependencies.

[0091] The progressive processing of alignment-weighting-interaction described above can solve the heterogeneity problem of multi-source features, and the key information is strengthened through dynamic weights and deep correlation mining. The final output fusion features can comprehensively reflect various factors affecting new energy power and their correlation, providing high-quality input for subsequent prediction.

[0092] In some embodiments, as shown in Figure 5 the method further includes:

[0093] S41, calculating a device state score based on the station type features, determining a matching volatility coefficient based on the equipment state features, calculating a station attribute evaluation value based on the device state score and the volatility coefficient, and matching the strength interval of the random feature based on the station attribute evaluation value;

[0094] S42, generating a random feature based on the strength interval, and inputting the random feature into the prediction model together with the fusion feature to obtain the predicted ultra-short-term new energy power of the target area.

[0095] The actual power generation output of a new energy station is not only affected by external factors such as weather and terrain, but also by random fluctuations caused by the characteristics of its own equipment, such as instantaneous power jumps that may occur in old photovoltaic components, unstable power generation caused by wear of fan gearboxes, and other irregular disturbances of the equipment itself that are difficult to fully model through multi-modal features. If the prediction model is directly input with fused features, it may ignore such random fluctuations, resulting in overly smoothed prediction results that cannot reflect the actual running deviations.

[0096] To address the above technical problems, the embodiment introduces a random feature based on a station attribute evaluation value to simulate the power generation output fluctuations caused by the instability of the equipment. For example, the higher the station attribute evaluation value (the worse the equipment state, the stronger the fluctuation), the stronger the random feature, so that the prediction model is based on the mapping relationship between the multi-modal features and the random disturbance and the actual power generation output, improves the prediction adaptability to irregular fluctuations, and ultimately reduces the deviation between the predicted value and the actual power generation output.

[0097] The random feature and the original fused feature are added element by element to form an enhanced fused feature, which not only retains the core correlation of multi-modal information, but also injects random disturbances matching the station attributes, can simulate the power generation output fluctuations caused by the instability of the equipment in actual operation, and improve the accuracy of predicting the target area of the ultra-short-term new energy power.

[0098] It can be understood that the prediction model is pre-trained based on the fused features fused with the random features, and the specific training process is as follows:

[0099] Step 1, select historical fused features, generate corresponding random features according to a determined random feature intensity interval, fuse the random features with the historical fused features to form a pre-training sample set containing random disturbances of different intensities;

[0100] Step 2, input the pre-training sample set into the initial prediction model, use the mean square error loss function, and iterate the training through the Adam optimizer until the loss of the model on the validation set converges; wherein, different intensity random features are randomly selected to inject into the fused features in each training round, so that the model learns the mapping rules in different equipment fluctuation scenarios;

[0101] Step 3, use the real-time collected fused features (containing corresponding random features) to fine-tune the pre-trained model parameters as initial values, and optimize the model by reducing the learning rate (such as reducing to 1 / 10 of the pre-training stage) to adapt to the latest equipment state and weather features of the current region.

[0102] In some embodiments, the device state score is calculated based on the station type feature, and the matching fluctuation coefficient is determined based on the device state feature, including:

[0103] The weight of each scoring dimension is determined based on the characteristics of the station type, and the device state score is calculated by synthesizing each quantitative indicator. The fluctuation coefficient is determined according to the device state characteristics and a preset matching relationship.

[0104] The device core influencing factors of different types of new energy stations are significantly different, so 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 photovoltaic, and agricultural-photovoltaic complementary, etc.).

[0105] For example: The device state of a centralized wind power station depends more on the "wind turbine blade wear degree" and "gearbox operating temperature", so the weights of these two quantitative indicators (such as 0.3 and 0.25) are higher than that of "cable aging degree" (weight 0.15); The distributed photovoltaic station is more affected by "component cleanliness" and "inverter conversion efficiency", so the corresponding indicator weights (such as 0.3 and 0.28) are significantly higher than those of other dimensions.

[0106] After determining the weights of each scoring dimension, the device state score is calculated by synthesizing each quantitative indicator (such as device operating life, fault repair time, performance decay rate, etc.) through weighted summation (range 0-1, the lower the score, the more stable the device state).

[0107] The device state characteristics (such as device service life, historical fault frequency, and the latest maintenance record, etc.) directly reflect its operating stability, and the fluctuation coefficient (range 0.1-0.8, the higher the coefficient, the greater the potential for device itself power output fluctuation) needs to be determined according to a preset matching relationship.

[0108] For example: In the preset rules, photovoltaic components with service life <5 years and no major faults correspond to a fluctuation coefficient of 0.2; components with service life ≥10 years and annual average faults ≥3 times correspond to a fluctuation coefficient of 0.6; for wind turbine equipment, gear box oil detection is qualified and there is no abnormal vibration in the last 3 months, the corresponding coefficient is 0.3; the bearing wear exceeds the standard and the historical fluctuation record is frequent, the corresponding coefficient is 0.7.

[0109] Through the above matching relationship, the qualitative description of the device state characteristics is converted into a quantitative coefficient to accurately describe the influence of the device itself characteristics on the power output fluctuation.

[0110] The above specific embodiments do not constitute a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A regional ultra-short-term renewable energy power prediction method based on multimodal information fusion, characterized in that, Includes the following steps: Acquire satellite remote sensing data, station observation data, meteorological model data, and station attribute data for the target area, and perform spatiotemporal alignment and outlier processing; Extract cloud system dynamics and surface radiation characteristics from satellite remote sensing data; extract power time series characteristics and local meteorological fluctuation characteristics from field observation data; extract regional meteorological trend characteristics from meteorological model data; and extract field type characteristics and equipment status characteristics from field attribute data. The extracted 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 are fused at multiple levels using a feature fusion network to obtain fused features. The number of fusion levels is determined based on the complexity index of data features, the dynamic index of the prediction scenario, and the feature correlation strength index. The fused features are input into the prediction model to obtain the predicted ultra-short-term renewable energy power of the target region; The number of fusion layers is determined at least based on data feature complexity indicators, prediction scenario dynamic indicators, and feature association strength indicators, including: 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 are calculated and used as the complex indicators of the data characteristics. Calculate the terrain complexity index, meteorological disaster warning level, and distribution density of new energy power stations in the target area, and use them as dynamic indicators for the predicted scenario; 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 are calculated and used as the feature correlation strength index. The data feature complexity index, the prediction scene dynamic index, and the feature association strength index are input into the attention weight network. The attention weights of each index are output through a fully connected layer and a softmax function, and the normalized score of each attention weight is calculated. The number of levels is determined based on the comprehensive score; wherein the comprehensive score is obtained by weighting each of the normalized scores.

2. The regional ultra-short-term renewable energy power prediction method based on multimodal information fusion according to claim 1, characterized in that: The feature fusion network includes a feature alignment layer, a weight adaptation layer, and a cross-modal interaction layer.

3. The regional ultra-short-term renewable energy power prediction method based on multimodal information fusion according to claim 2, characterized in that: A feature fusion network is used to fuse the extracted 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 at multiple levels 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 to a unified high-dimensional feature space. It achieves alignment of time step and spatial scale through spatiotemporal coordinate transformation and outputs aligned features with consistent dimensions. The weighted adaptive layer assigns dynamic weights to each type of alignment feature after alignment, and then concatenates the alignment features of each type based on the dynamic weights 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 explores the potential correlations between different features, performs feature fusion based on the potential correlations, and outputs the integrated fused features.

4. The regional ultra-short-term renewable energy power prediction method based on multimodal information fusion according to claim 3, characterized in that: The weight values ​​of the dynamic weights are dynamically adjusted based on the feature association strength index, and the dynamic weights of cloud dynamic features and surface radiation features, power time series features and local meteorological fluctuation features are strengthened through an attention mechanism.

5. The regional ultra-short-term renewable energy power prediction method based on multimodal information fusion according to claim 1, characterized in that: The fused features are input into the prediction model to obtain the predicted ultra-short-term renewable energy power of the target region, including: Based on the characteristics of the site type, an equipment status score is calculated, a matching volatility coefficient is determined based on the equipment status characteristics, a site attribute evaluation value is calculated based on the equipment status score and the volatility coefficient, and the intensity range of random features is matched based on the site attribute evaluation value. Random features are generated based on the intensity range, and these random features are then incorporated into the fused features and input together into the prediction model to obtain the predicted ultra-short-term renewable energy power of the target region.

6. The regional ultra-short-term renewable energy power prediction method based on multimodal information fusion according to claim 5, characterized in that: The process of calculating equipment status scores based on site type characteristics and determining matching volatility coefficients based on equipment status characteristics includes: The weights of each scoring dimension are determined based on the characteristics of the site type, and the equipment status score is calculated by combining various quantitative indicators; the volatility coefficient is determined based on the equipment status characteristics and the preset matching relationship.

Images