Deep learning-based multi-temporal and multi-spatial scale power meteorological data fusion analysis method

Abstract

The application relates to the technical field of safe operation and data analysis processing of power systems, and discloses a multi-time-space-scale power meteorological data fusion analysis method based on deep learning, which comprises the following steps: S1, collecting and fusing multi-source heterogeneous data of sky layers, ground layers and network layers, and performing data preprocessing to construct a unified power meteorological feature data set; S2, based on the power meteorological feature data set, multi-time-scale features and multi-space-scale features are respectively constructed; S3, a prediction model based on a time-aware multi-scale recurrent neural network is constructed and trained; and S4, feature data of a region to be predicted is input into the trained prediction model, and prediction results of various power meteorological elements in a specified period in the future are output. Through the construction of the multi-time-space-scale deep learning prediction model, high-resolution, fine and customized prediction of key power meteorological elements is realized.

Description

Technical Field

[0001] This invention relates to the field of power system safe operation and data analysis and processing technology, specifically to a multi-temporal and spatial scale power meteorological data fusion and analysis method based on deep learning. Background Technology

[0002] With the increasing penetration of new energy sources and the transformation of the power system towards low-carbon development, meteorological conditions have a more significant impact on the safe and stable operation of the power system, especially in regions with complex terrain and diverse weather systems, such as Northwest China. This region is influenced by the superposition of the westerly winds and the subtropical circulation, resulting in the interplay of large-scale circulation systems (such as the Ural trough and the plateau trough) and mesoscale systems (such as the low-level jet stream and shear lines). This easily triggers severe weather events such as strong winds, sandstorms, blizzards, and icing. Simultaneously, the steep terrain and large elevation differences lead to prominent local micro-meteorological characteristics, posing serious challenges to the safety of power facilities, the prediction of new energy power output, and grid dispatching.

[0003] Current power meteorological forecasting methods mainly rely on numerical weather prediction models and traditional machine learning algorithms, but they have significant drawbacks: First, numerical weather prediction models are poorly adaptable to complex terrain and have low spatial resolution (mostly at the kilometer level), making it difficult to meet the data fusion and analysis needs at the equipment scale (100-meter level); Second, traditional machine learning algorithms (such as random forests and support vector machines) require manually designed features and cannot effectively capture the multi-temporal and spatial scale dependencies of power meteorological data; Third, existing deep learning models (such as basic RNNs and LSTMs) mostly use fixed-scale modeling and do not consider the dynamic evolution characteristics of meteorological elements at different time (minute-level sudden weather, daily trends, seasonal circulation) and spatial (regional, station, equipment) scales, resulting in insufficient prediction accuracy; Fourth, modeling for multi-scale features often suffers from feature entanglement problems, making it impossible to achieve independent learning and dynamic adaptation of information at different scales, and making it difficult to match the differentiated needs of different forecast scenarios such as ultra-short-term, short-term, and medium-term forecasts.

[0004] In addition, existing methods do not fully integrate multi-source heterogeneous data from "space-ground-grid" (satellite remote sensing, ground monitoring, and power grid operation data), and lack targeted modeling for the unique hazardous weather in Northwest China (such as sandstorms and strong winds caused by canyon funneling effects). This results in poor adaptability of forecast results to power business, making it difficult to support actual needs such as real-time power grid dispatch, renewable energy consumption, and equipment maintenance planning. Summary of the Invention

[0005] To address the problems of low accuracy, poor multi-scale adaptability, and insufficient operational relevance of existing power meteorological forecasting methods in complex terrain areas, this invention proposes a deep learning-based multi-temporal and spatial scale power meteorological data fusion and analysis method. Addressing the safety operation needs of the power system in Northwest China, this method integrates multi-source power meteorological observation data, numerical weather prediction data, and grid operation correlation information to construct a multi-temporal and spatial scale deep learning prediction model. This enables high-resolution, refined, and customized forecasts of key power meteorological elements, and is particularly suitable for power meteorological element prediction in complex terrain areas (such as the area where the three major plateaus of Northwest China converge). It can be widely applied to scenarios such as transmission line safety monitoring, new energy power plant power prediction, grid dispatch optimization, and power operation safety risk assessment.

[0006] The following technical solution is used to achieve the following: a deep learning-based multi-temporal and spatial scale power meteorological data fusion and analysis method, comprising the following steps: S1, collecting and fusing multi-source heterogeneous data from the sky layer, ground layer, and network layer, and preprocessing the data to construct a unified power meteorological feature dataset; S2, based on the power meteorological feature dataset, constructing multi-temporal scale features adapted to the ultra-short-term, short-term, and medium-term forecasting needs of power operations, as well as multi-spatial scale features covering the area, station, and equipment scales; S3, constructing and training a prediction model based on a time-aware multi-scale recurrent neural network, wherein the prediction model is used to perform fusion analysis on the multi-temporal scale features and the multi-spatial scale features. The hidden states formed by the loop unit during the processing of time-series data are decoupled into multiple parallel small hidden states used to characterize the evolution characteristics at different time scales; each small hidden state is configured with an independent update mechanism with a differentiated update frequency; based on the current time-series input information and historical context, the importance weight of each small hidden state to the current prediction is dynamically determined, and the corresponding historical small hidden state information is modulated according to the importance weight; based on the modulated small hidden states, the final power meteorological element prediction results are generated; S4, the feature data of the area to be predicted is input into the prediction model trained in step S3, and the prediction results of multiple power meteorological elements in the future specified time period are output.

[0007] Furthermore, in step S1, the sky layer data includes meteorological satellite remote sensing data and weather radar data; the ground layer data includes meteorological station observation data distributed at different terrains and altitudes, new energy power station monitoring data, transmission line micro-meteorological monitoring data, and special monitoring data for severe weather; the network layer data includes power grid operation status data, power equipment fault records, power operation safety data, and power grid geographic information data.

[0008] Furthermore, in step S3, the prediction model includes an input layer, a multi-scale feature decoupling layer, a time-aware feature modulation layer, a fusion layer, and an output layer connected in sequence; the input layer is used to receive an input feature sequence composed of multi-time-scale features and multi-spatial-scale features; the multi-scale feature decoupling layer is used to split the hidden state of the previous time step into multiple independent small hidden states, and control each small hidden state to selectively update according to its preset update frequency, wherein the selective update includes keeping the state unchanged at non-update times, and calculating the state based on the current input and the historical value corresponding to the small hidden state at update times; the time-aware feature decoupling layer... The feature modulation layer is used to calculate the dynamic weights corresponding to each small hidden state based on the current input features and the complete hidden state of the previous time step. When the multi-scale feature decoupling layer performs the update calculation, it first uses the corresponding scale importance weight to adjust the historical value of the small hidden state, and then combines the adjusted historical value with the current input features to calculate the new small hidden state value at the current time through nonlinear transformation. The fusion layer is used to fuse the multiple modulated small hidden states to form a unified global feature representation. The output layer is used to map the global feature representation into the predicted values ​​of various power meteorological elements.

[0009] Furthermore, the specific operations of the multi-scale feature decoupling layer include: [decoupling the hidden state at time step t]. Split into K small hidden states k=1,2,…,K, each small hidden state Corresponding to an update scale For the k-th small hidden state at time step t Its update rule is: if the current time step t is not the update scale... If the value is an integer multiple of the previous time step, then the current state copies the state from the previous time step. Otherwise, update based on the input feature vector at time t and the state at the previous time step, i.e. ;in, Let be the input feature vector at time t. For the input weight matrix, For a recursive matrix, Here, tanh is the bias term, and tanh is the hyperbolic tangent activation function. This represents the k-th small hidden state at the (t-1)-th time step.

[0010] Furthermore, the number of small hidden states K=4, and the update scale set S is... ; It is used to capture time scale dependencies at the minute / hour, hour / day, day / week, and week / seasonal levels.

[0011] Furthermore, the time-aware feature modulation layer calculates the dynamic weights through a weight generation unit; the weight generation unit receives the current input feature vector and the complete hidden state vector of the previous time step, performs linear transformation and normalization processing, and outputs a set of weight values ​​that sum to 1, with each weight value corresponding to a small hidden state.

[0012] Furthermore, in step S4, the prediction model is trained using a hybrid loss function L, which is a weighted sum of the mean squared error (MSE) and the mean absolute percentage error (MAPE). ;in, For the true value, For predicted values, and These are the weighting coefficients.

[0013] Furthermore, in step S4, the power meteorological elements include at least two of the following: wind speed, wind direction, gust intensity, temperature, humidity, air pressure, precipitation intensity, snow depth, icing thickness, total irradiance, dust concentration, and thunderstorm probability.

[0014] Furthermore, step S4 specifically includes: in response to ultra-short-term forecast scenarios, the prediction model inputs a high-frequency feature sequence of the past 1 hour or 6 hours and outputs a prediction result for the next 0-6 hours with a time resolution of minutes; in response to short-term forecast scenarios, the prediction model inputs a feature sequence of the past 72 hours and outputs a prediction result for the next 1-3 days with a time resolution of hours; in response to medium-term forecast scenarios, the prediction model inputs a daily average feature sequence of the past 10 days and outputs a prediction result for the next 3-10 days with a time resolution of days.

[0015] Furthermore, in step S2, the multi-spatial-scale features include: regional-scale features covering a province or river basin, station-scale features covering new energy power plants, and equipment-scale features covering transmission lines or substations; in step S1, data preprocessing includes spatial interpolation, which is used to downscale regional-scale meteorological data to the equipment scale to match the spatial distribution of power grid equipment, wherein the spatial resolution of the equipment scale is 100 meters.

[0016] Compared with the prior art, the present invention achieves the following beneficial effects:

[0017] 1. Outstanding multi-scale modeling capabilities and significantly improved prediction accuracy: This invention uses Time-Aware Multi-Scale Recurrent Neural Networks (TAMS-RNNs) as the core prediction model. Through feature decoupling and time-aware modulation mechanisms, it adaptively captures the multi-temporal and spatial scale dependencies of power meteorological data, achieving collaborative learning of features at different time and spatial scales and improving the prediction accuracy of meteorological elements in complex terrain areas. Specifically, the hidden state is decoupled into four independent small hidden states through a Multi-Scale Feature Disentanglement (MSFD) layer, each capturing dependencies at different time scales. Combined with a Time-Aware Feature Modulation (TAFM) layer that dynamically modulates the scale weights, the model can accurately adapt to the multi-scale evolution patterns of complex meteorological elements.

[0018] 2. Strong adaptability to complex terrain and business scenarios: In response to the characteristics of the three plateaus in Northwest China and the steep terrain, the system integrates multi-source data such as satellite remote sensing, ground micro-meteorological monitoring, and power grid operation to construct three-level spatial scale features of "region-station-equipment" and three-level temporal scale features of "ultra-short-term-short-term-medium-term". The output resolution reaches 100 meters (spatial) and minutes (time), which can accurately match the needs of different business scenarios such as transmission line icing monitoring, wind farm wind speed prediction, and power operation safety early warning.

[0019] 3. Excellent model interpretability and robustness: The Multi-Scale Feature Decoupling Layer (MSFD) enables independent decoupling of multi-scale features, and the dynamic weights of the Time-Aware Feature Modulation Layer (TAFM) can intuitively reflect the contribution of different scales to the forecast results, improving the model interpretability. At the same time, the model effectively suppresses overfitting through Dropout regularization, weight decay, and electrical physical constraint correction, and has strong tolerance to noise such as cloud cover. It can still maintain stable prediction performance in complex scenarios with frequent severe weather.

[0020] 4. Significant business value and wide range of applications: This invention can simultaneously output 12 key power meteorological elements such as wind speed, wind direction, icing thickness, irradiance, and thunderstorm probability. Ultra-short-term forecasts support real-time grid dispatch and operational safety decisions, short-term forecasts assist in new energy power prediction and grid operation planning, and medium-term forecasts provide a basis for equipment maintenance plans and risk assessments. It comprehensively covers the meteorological support needs of the entire process of power system "monitoring-forecasting-dispatch-operation and maintenance", effectively reducing the risk of grid outages caused by severe weather, and improving the capacity for new energy absorption and the economic efficiency of grid operation. Attached Figure Description

[0021] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0022] Figure 1 This is a flowchart illustrating the multi-temporal-spatial-scale power meteorological data fusion and analysis method based on deep learning provided in this embodiment of the invention.

[0023] Figure 2 This is a schematic diagram of the system architecture according to an embodiment of the present invention;

[0024] Figure 3 This is a schematic diagram of the prediction model in an embodiment of the present invention;

[0025] Figure 4 This is a schematic diagram of the weather forecast visualization results according to an embodiment of the present invention;

[0026] Figure 5 This is a comparison chart of the accuracy of different models provided in the embodiments of the present invention. Detailed Implementation

[0027] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0028] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0029] The acquisition, storage, use, and processing of data in this application comply with relevant national laws and regulations. It should be noted that existing industry solutions such as software, components, or models may be mentioned in the embodiments of this application. These should be considered exemplary and intended only to illustrate the feasibility of implementing the technical solution of this application, but do not imply that the applicant has already used or necessarily used such a solution.

[0030] Figure 1This is a flowchart illustrating the multi-temporal-spatial-scale power meteorological data fusion and analysis method based on deep learning provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the system architecture of the multi-temporal and spatial scale power and meteorological data fusion and analysis method based on deep learning, according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the deep learning-based multi-temporal-scale power meteorological data fusion and analysis method includes the following steps:

[0031] S1. Collect and fuse multi-source heterogeneous data from the sky layer, ground layer, and network layer, and construct a unified power meteorological feature dataset after data preprocessing;

[0032] Step S1 is used to perform data acquisition and preprocessing, including the following steps:

[0033] S1.1: Integrated "Sky-Ground-Network" power meteorological data acquisition combines the core characteristics of Northwest China—"the convergence of three major plateaus (Qinghai-Tibet Plateau, Loess Plateau, and Inner Mongolia Plateau), steep terrain with large elevation differences, complex weather systems (overlapping westerly winds and subtropical circulation, intertwined large-scale and mesoscale systems), and frequent severe weather"—with the actual needs of safe operation of power facilities and safe power work. It collects multi-source heterogeneous data covering the sky, ground, and network dimensions to ensure that the data accurately depicts regional meteorological characteristics and is adaptable to power operations, as detailed below:

[0034] S1.1.1: Collect sky layer data

[0035] Focusing on monitoring the dominant weather systems in Northwest China, high spatiotemporal resolution data were collected to capture large-scale circulation and mesoscale convection characteristics. Sky layer data included meteorological satellite remote sensing data and weather radar data.

[0036] (1) Meteorological satellite remote sensing data: FY series and Sentinel series satellite data were selected, including cloud top height, cloud cover, atmospheric circulation parameters (suitable for monitoring large-scale systems such as the Ural trough and the plateau trough), dust concentration (for dust storm disasters), and snow cover range (suitable for blizzard disasters). The temporal resolution is 15 minutes and the spatial resolution is 1 km, covering the westerly wind belt and the subtropical circulation area.

[0037] (2) Radar reflectivity data: The new generation of Doppler weather radar network data in the Northwest region is used to capture strong convection, rainstorms and thunderstorms caused by mesoscale systems (low-level jet stream, shear line and squall line). The time resolution is 6 minutes and the spatial resolution is 0.5 km, which is suitable for monitoring the rapid evolution of local severe weather under complex terrain.

[0038] S1.1.2: Collect ground layer data

[0039] In response to the characteristics of the Northwest region, which features large topographical differences and concentrated disaster types, the density of monitoring points in key areas was increased to collect targeted meteorological and equipment-related data. Surface-level data includes observation data from meteorological stations distributed across different terrains and altitudes, monitoring data from new energy power plants, micro-meteorological monitoring data from transmission lines, and special monitoring data for severe weather. Specifically:

[0040] (1) Conventional meteorological station data: Meteorological station observation data covering different altitudes (from plains to plateaus) and terrain types (canyons, mountains, plateaus), including temperature, humidity, air pressure, precipitation intensity, wind speed, and wind direction (adapted to the wind system switching characteristics of northerly / westerly winds in winter and southerly / easterly winds in summer), updated hourly, with a focus on supplementing low temperature (dry and cold weather in winter) and snow depth (blizzard disasters);

[0041] (2) Special data for new energy power stations: wind farm wind measurement tower data (covering different heights of 10m / 50m / 100m, adapting to the differences in vertical distribution of wind speed in plateau canyons), photovoltaic irradiance monitoring data (total irradiance, diffuse irradiance, adapting to the strong irradiance characteristics of plateau), updated in minutes, supporting the prediction of new energy power.

[0042] (3) Micro-meteorological monitoring data of transmission lines: In the sections of the line with steep terrain, prone to icing, road icing, and landslides, micro-meteorological monitoring devices are deployed to collect data on conductor temperature, icing thickness, local wind speed (adapting to the canyon terrain and narrow tube effect), surface temperature (low temperature monitoring), and soil moisture (early warning of secondary disasters such as landslides and debris flows), with updates every minute;

[0043] (4) Special monitoring of severe weather: Add dust concentration monitoring equipment in areas with high incidence of dust storms, supplement soil moisture monitoring in areas with frequent rainstorms, and increase the number of automatic snow depth observation points in areas affected by snowstorms to meet the monitoring needs of special disasters and secondary disasters in Northwest China.

[0044] S1.1.3: Collect network layer data

[0045] Focusing on the pain points of power production in Northwest China, this study collects power business data strongly correlated with meteorological conditions to compensate for the shortcomings of existing numerical models that are not adapted to power demand. Network layer data includes power grid operation status data, power equipment fault records, power operation safety data, and power grid geographic information data. Specifically:

[0046] (1) Power grid operation monitoring data: covering active / reactive power of wind farms and photovoltaic power plants (adapting to the demand for new energy power forecasting), transmission line load rate, substation operation status, updated in minutes, capturing the real-time impact of severe weather on power grid operation;

[0047] (2) Power equipment status and fault data: historical equipment fault records (including line galloping caused by strong winds, equipment short circuits caused by sandstorms, equipment tripping due to icing, and equipment shutdowns caused by blizzards), equipment defect data (related to low temperature and sand deposition), and associated with meteorological conditions of the corresponding time period to construct a "weather-fault" mapping relationship;

[0048] (3) Power operation safety data: real-time weather records during power operation (such as wind speed, precipitation, visibility, and dust concentration during operation), data on the correlation between operation interruption / safety accidents and meteorological factors, to support early warning of operation safety risks;

[0049] (4) Topographic and power grid correlation data: The precise geographic coordinates, altitude, terrain type (canyon, plateau, mountain) and surrounding terrain obstruction of power grid equipment (lines, stations, substations) provide basic data for subsequent spatial interpolation and terrain correction.

[0050] S1.2: Multi-source data preprocessing

[0051] The collected heterogeneous data were standardized to form a unified format power meteorological feature dataset:

[0052] (1) Anomaly removal: Based on the 3σ criterion, outliers that exceed the physical reasonable range are removed (such as wind speed > 70m / s, air temperature > 60℃, etc.).

[0053] (2) Time synchronization: All data are uniformly resampled to minute-level (ultra-short-term forecast), hour-level (short-term forecast), and day-level (medium-term forecast) time scales to ensure consistent timestamps;

[0054] (3) Spatial interpolation: Kriging interpolation is used to downscale regional meteorological data to the equipment scale (100-meter resolution) to match the spatial distribution of power grid equipment;

[0055] (4) Standardization: Perform Z-score standardization on all features to eliminate dimensional differences.

[0056] S2. Based on the power meteorological feature dataset, multi-timescale features adapted to the ultra-short-term, short-term and medium-term forecasting needs of power business are constructed, as well as multi-spatial-scale features covering the area, station and equipment scales.

[0057] Step S2 is used to construct multi-temporal scale features, which specifically includes the following steps:

[0058] S2.1: Multi-timescale feature modeling

[0059] To address the varying forecast lead times required for power services, a three-tiered timescale input feature model is constructed to adapt to ultra-short-term, short-term, and medium-term forecasting scenarios.

[0060] (1) Ultra-short-term scale (minute level, hour level): Select high-frequency data sequences of the last 1 hour / 6 hours, including rapid evolution characteristics such as wind speed, wind direction, and precipitation intensity, for use in forecasting sudden weather events such as line icing and thunderstorms;

[0061] (2) Short-term scale (1-3 days): Select the data series of the past 72 hours, including the daily temperature variation, humidity trend, and movement of the air pressure system, for power grid dispatch and new energy power prediction;

[0062] (3) Medium-term scale (3-10 days): Select the daily average data series of the past 10 days, including seasonal trends, atmospheric circulation background and other characteristics, for maintenance planning and risk assessment.

[0063] S2.2: Multi-scale feature modeling

[0064] By combining complex terrain and the spatial distribution of the power grid, a three-level spatial scale feature representation is constructed to achieve refined coverage from region to equipment:

[0065] (1) Regional scale (province, watershed): extract meteorological field features of 5km×5km grid, including large-scale circulation features such as pressure field, temperature field, and wind field;

[0066] (2) Site scale (wind farm, photovoltaic base): Extract local meteorological features of 1km×1km grid, including the spatial distribution features of wind speed and irradiance after terrain correction;

[0067] (3) Equipment scale (transmission lines, substations): extract micro-meteorological features of a 100m×100m grid, including local wind speed around the conductor, terrain shading coefficient, icing sensitivity factor, etc.

[0068] Through steps S1 and S2, this invention integrates multi-source data from the "sky-ground-grid" system to construct a multi-temporal and spatial scale feature system. Combined with the physical constraints of power operations, it achieves data fusion analysis and refined forecasting of key power meteorological elements at the equipment scale (100-meter level) and minute-level resolution, providing customized meteorological support for the safe operation of the power system.

[0069] S3. Construct and train a prediction model based on a time-aware multi-scale recurrent neural network, wherein the prediction model is used to perform the following operations: decouple the hidden states formed by the recurrent unit during the processing of time-series data into multiple parallel small hidden states that characterize the evolution features at different time scales; configure an independent update mechanism with differentiated update frequencies for each small hidden state; dynamically determine the importance weight of each small hidden state to the current prediction based on the current time-series input information and historical context, and modulate the corresponding historical small hidden state information according to the importance weight; generate the final prediction result of power meteorological elements based on the modulated small hidden states.

[0070] Step S3 is used to construct a deep learning prediction model based on Time-Aware Multi-Scale Recurrent Neural Networks (TAMS-RNNs). The TAMS-RNN prediction model proposed in this embodiment of the invention adaptively captures multi-spatiotemporal scale dependencies through feature decoupling and time-aware modulation, as specifically designed below:

[0071] S3.1: Model Structure Design

[0072] like Figure 3 The diagram shown is a schematic representation of the prediction model in an embodiment of the present invention. The prediction model based on Time-Aware Multi-Scale Recurrent Neural Networks (TAMS-RNNs) consists of an input layer, a Multi-Scale Feature Disentanglement (MSFD) layer, a Time-Aware Feature Modulation (TAFM) layer, a hidden state fusion layer, an output layer, and a constraint correction layer, as shown below:

[0073] S3.1.1: Constructing the Input Layer

[0074] The input layer receives an input feature vector composed of features from multiple temporal and spatial scales. The input features are multi-temporal and spatial scale fused feature vectors. Dimension is defined as ,in For time steps (ultra-short-term forecasts) (24 hours) Minute-time and short-term forecasts (72 hours) Hourly and medium-term forecasts (10 days) sky)), The feature dimensions include 42 features after multi-source data fusion, including three sub-features: meteorology, power, and topography.

[0075] X: Multi-temporal and spatial scale fusion feature vector, which is the core input data of the model input layer; : The dimensional space of the input feature vector X, representing the time step and feature dimension attribute of the vector; T: The time step of the input feature sequence, for ultra-short-term forecasting. (24 hours) Minute-time and short-term forecasts (72 hours) Hourly and medium-term forecasts (10 days) D: The dimension of the input feature, which is a 42-dimensional feature after the fusion of multi-source data (including three sub-features: meteorology, electricity, and topography).

[0076] S3.1.2: Constructing a Multi-Scale Feature Decoupling Layer (MSFD)

[0077] A single hidden state is decoupled into K independently updated smaller hidden states to achieve feature separation across multiple time scales. Specifically, the hidden state of the previous time step is split into multiple independent smaller hidden states, and each smaller hidden state is selectively updated according to its preset update frequency. Selective updates include keeping the state unchanged at non-update times and calculating the state based on the current input and the historical values ​​corresponding to the smaller hidden state at update times. More specifically, the following operations are performed:

[0078] 1. Hidden State Decomposition: The hidden state of the recurrent unit is decomposed into multiple parallel smaller hidden states, each used to independently capture features at a specific time scale. As a preferred embodiment, the hidden state of the RNN at time step t is... Split into A small hidden state Each hidden state has a dimension of 64. Each hidden state is equipped with an independent recursion matrix. and input weight matrix .

[0079] RNN in the 19th century A single hidden state at each time step is the basic hidden state before multi-scale feature decoupling. Single hidden state The dimensional space has 256 dimensions; The number of small hidden states after splitting is preferably K, which is a fixed value of 4, corresponding to 4 different time scales; : No. The time step A small hidden state It is used to independently capture features at different time scales; Single small hidden state The dimension space is 64-dimensional (256 total hidden states are evenly distributed into 4 smaller hidden states). : No. The independent recursive matrices corresponding to each small hidden state are used to model the temporal dependency of the hidden states at this scale. Recursive matrix The dimension space has 64 rows (matching the small hidden state dimension) and 64 columns (matching the small hidden state dimension of the previous time step). : No. The independent input weight matrix corresponding to each small hidden state is used to map the input features to the hidden state space of that scale; Input weight matrix The dimension space has 64 rows (matching the small hidden state dimension) and 42 columns (matching the input feature dimension D).

[0080] 2. Differentiated update frequency: Assign scale sets to the four small hidden states. = Each small hidden state Corresponding to an update scale Used to capture timescale dependencies at the minute / hour, hour / day, day / week, and week / seasonal levels:

[0081] The scale set for the four small hidden states, corresponding to different time step update intervals, takes values ​​of... This corresponds to different time step update intervals; The update scale of the first small hidden state, with a value of 1, means it is updated once every time step, and is used to capture short-term dependencies at the minute / hour level (such as sudden gusts of wind). The update scale of the second small hidden state, with a value of 2, means it is updated once every 2 time steps. It is used to capture hourly / daily medium-term dependencies (such as daily temperature changes). The update scale of the third hidden state is set to 4, meaning it is updated every 4 time steps. This is used to capture long-term dependencies at the daily / weekly level (such as the movement of circulation systems). The update scale of the 4th small hidden state is set to 8, meaning it is updated every 8 time steps. This is used to capture weekly / seasonal dependencies (such as seasonal wind trends).

[0082] 3. Independent Update Mechanism: For the kth hidden state, the update rule is as follows:

[0083]

[0084] The time step index represents the current time step in the calculation and is a core index parameter for time series feature modeling. : No. The time step for the first Individual scale The modulo operation is used to determine whether the current time step is the update step of the small hidden state; : No. The time step The small hidden state, i.e. the small hidden state of the previous time step, provides historical time information for the current step update; tanh: hyperbolic tangent activation function, used to realize non-linear feature transformation when updating the small hidden state, and improve the model's ability to fit complex time features. : No. The input feature vector at each time step is the single-step input data passed from the input layer to the multi-scale feature decoupling layer; : Bias term during small hidden state update, used to adjust the baseline value of hidden state update and optimize feature mapping effect; Bias term The dimension space is 64-dimensional (matching the small hidden state dimension); copy: when the... The time step for the first The update scale of a small hidden state If the modulo result is not 0, it means that the current time step is not the "update period" of the small hidden state at this scale. In this case, the small hidden state of the previous time step is directly updated. Assigned to the current time step (Right now Update: When the first The time step for the first The update scale of a small hidden state When the modulo result is 0, it indicates that the current time step is the "update period" of the small-scale hidden state. At this point, [the process is as follows]. Calculate the new hidden state value, complete. Update.

[0085] This independent update mechanism indicates that if the current time step t is not an update scale... Integer multiples (i.e.) If the current state copies the state from the previous time step, then the current state copies the state from the previous time step. Otherwise, update based on the current input and the previous state, i.e. .

[0086] S3.1.3: Constructing a Time-Aware Feature Modulation Layer (TAFM)

[0087] The weights at each scale are dynamically adjusted based on the temporal context to adapt to the scale requirements of different forecast scenarios. Specifically, the time-aware feature modulation layer calculates the dynamic weights corresponding to each small hidden state based on the current input features and the complete hidden state of the previous time step. When the multi-scale feature decoupling layer performs update calculations, it first uses the corresponding scale importance weight to weight and adjust the historical value of the small hidden state. Then, it combines the adjusted historical value with the current input features and calculates the new small hidden state value at the current time step through a nonlinear transformation. More specifically, the following operations are performed:

[0088] 1. Weight Generation: The Time-Aware Feature Modulation (TAFM) layer calculates dynamic weights through a weight generation unit; the weight generation unit receives the current input feature vector. The complete hidden state vector from the previous time step A set of scale importance weights summing to 1 is generated through linear transformation and SoftMax. Each scale importance weight value corresponds to a small hidden state, and the specific calculation formula is as follows:

[0089]

[0090] : The scale importance weight vector generated by the time-aware feature modulation layer is used to dynamically adjust the importance of the four small hidden states to adapt to the needs of different forecast scenarios; softmax: The softmax activation function normalizes the original weight values ​​after linear transformation to the [0,1] interval to ensure that the sum of the weights of each scale is 1, which is convenient for quantifying the relative importance of different scales. : The learnable weight matrix corresponding to the current input features, used to weight the input features at time step t. Mapped to scale weight space; Weight matrix The dimension space has 4 rows (corresponding to 4 small hidden states / scales) and 42 columns (matching the input feature dimension D). The input feature vector at time step t contains multi-source fused power meteorological and terrain features, providing current time-series information for weight generation; : The learnable weight matrix corresponding to the hidden state in the previous step, used to map the global hidden state h_(t-1) at time step t-1 to the scale weight space; Weight matrix The dimensional space has 4 rows (corresponding to 4 scales) and 256 columns (matching the global hidden state). (dimensions) The global hidden state at time step t-1 (composed of 4 smaller hidden states) provides historical temporal context information for weight generation. : Learnable bias terms in the scale weight generation process, used to adjust the baseline offset of the weights and optimize the rationality of weight allocation; Bias term The dimensional space has 4 dimensions (corresponding to 4 scales). Weight vector The k-th element (k=1,2,3,4) represents the importance of the k-th scale, and its value ranges from [0,1], satisfying the following condition: .

[0091] 2. Dynamic Modulation: Only in the "Update" step, using weights Modulate the small hidden state to avoid gradient vanishing:

[0092] (Updating)

[0093] : The updated value of the kth small hidden state at the t-th time step after dynamic modulation (only applicable to the "update" step); tanh: Hyperbolic tangent activation function, used to perform nonlinear transformation on the modulated features to improve the model's ability to fit complex power meteorological time series features; The independent input weight matrix corresponding to the k-th small hidden state, and the multi-scale feature decoupling layer (MSFD) in... Consistency, used to integrate input features Mapped to the hidden state space at this scale; The independent recursive matrix corresponding to the k-th small hidden state, and the matrix in the multi-scale feature decoupling layer (MSFD). Consistency is used to model temporal dependencies at this scale; The weight of the k-th scale is combined with the Hadamard product (element-wise product) of the small hidden state of the previous time step to achieve dynamic modulation of the historical state at that scale. The larger the weight, the stronger the influence of the historical state on the current update. The k-th small hidden state at time step t-1 provides historical temporal information at this scale for the current step update; Learnable bias terms during small hidden state updates, and in the multi-scale feature decoupling layer (MSFD) Consistency is used to adjust the baseline value for hidden state updates and optimize feature mapping performance.

[0094] S3.1.4: Constructing the Hidden State Fusion and Output Layer

[0095] 1. Fusion Layer: This layer integrates the modulated multi-scale small hidden states into a network layer with unified global features. It contains one fully connected layer, specifically concatenating the four modulated small hidden states into a complete hidden state. It is mapped to a high-dimensional feature vector through a single fully connected layer (with ReLU activation function). High-dimensional feature vectors It is the feature representation output by the fusion layer, which is the core feature after multi-scale feature integration and nonlinear transformation.

[0096] After time-aware feature modulation, the four independent small hidden states at the t-th time step correspond to the feature representations at four different time scales. The complete global hidden state is formed by concatenating the four smaller hidden states along the dimensions, and multi-scale temporal features are integrated. Complete global hidden state The dimension space is composed of four 64-dimensional small hidden states (4×64=256); ReLU: The linear rectified activation function used in the fusion layer and the output layer can avoid gradient vanishing and adapt to the non-negativity characteristics of power meteorological elements such as wind speed and icing thickness.

[0097] 2. Output Layer: The output layer of TAMS-RNNs is a multi-dimensional regression layer, which represents the high-dimensional features after fusing the hidden states, i.e., the global features, and simultaneously maps them to the predicted values ​​of 12 types of power meteorological elements. The formula is defined as follows:

[0098]

[0099] The output feature matrix of the hidden state fusion layer is the core input data after integrating multi-scale features; : Feature matrix dimensional space, This corresponds to the output temporal length of the forecast scene, where 256 represents the dimension of the fused features; the output temporal length corresponding to the forecast scene. Determined by the scenario type (ultra-short-term forecast) Short-term forecast Medium-term forecast ; Output layer fully connected weight matrix, used to map high-dimensional fused features into predicted values ​​of 12 types of power meteorological elements; Weight matrix The dimensional space, with 256 rows (matching) (Feature dimensions), number of columns 12 (corresponding to 12 categories of key power meteorological elements); : Output layer bias vector, used to adjust the baseline offset of the predicted values ​​of 12 types of power meteorological elements and optimize the prediction accuracy; Bias vector The dimensional space has 12 dimensions (corresponding to 12 types of power meteorological elements); : Contextualized original prediction output matrix The dimension space is denoted by , with each row corresponding to the predicted values ​​of 12 types of elements at one time step, and each column corresponding to one specific element; ReLU: activation function, to avoid gradient vanishing and adapt to the non-negative characteristics of meteorological elements (such as wind speed, ice thickness, etc.). : Scenario-based raw prediction output matrix, each row corresponds to the predicted value of 12 categories of features at one time step, and the columns correspond to specific features and time series lengths. It is determined by the forecast scenario (as shown in Table 1).

[0100] Table 1

[0101]

[0102] S3.2: Model Training Steps and Key Parameters

[0103] S3.2.1: Source of Training Data

[0104] The training data uses integrated "space-ground-network" data from the past 5 years, divided into training, validation, and test sets in a 7:1:2 ratio.

[0105] (1) Training set: used for updating model parameters;

[0106] (2) Validation set: used for hyperparameter tuning and early stopping judgment;

[0107] (3) Test set: used for model performance evaluation.

[0108] S3.2.2: Training Parameter Settings

[0109] (1) Number of network layers: TAMS-RNNs is a 2-layer stacked structure with 256 hidden state dimensions in each layer; the hidden state dimension is the feature dimension of the hidden state in each layer of the network, which is used to carry multi-temporal and spatial scale power meteorological feature information.

[0110] (2) Optimizer: The parameter optimization algorithm used for model training is the Adam optimizer, which is adapted to the gradient descent optimization requirements of deep learning models. The initial learning rate is 0.001. The learning rate decay rule, i.e., the dynamic adjustment strategy of the learning rate, is: after every 10 epochs, the learning rate decays to 0.9 of the current value to avoid parameter oscillation in the later stage of training and improve convergence accuracy. Among them, the initial learning rate is the initial learning rate of the Adam optimizer, which is 0.001, to control the step size of the initial parameter update and ensure training stability; epoch is the iteration unit of model training, and 1 epoch means that the training set data is completely traversed once.

[0111] (3) Regularization: A combination of techniques used to suppress model overfitting, including Dropout layer and weight decay, to ensure the model's generalization ability. Specifically: A Dropout layer (dropout rate 0.1) is added to the input layer, and the weight decay coefficient λ = 1e-4 is used to prevent overfitting. The Dropout layer is a regularization component added to the input layer, which simulates data augmentation by randomly deactivating some features; the dropout rate is the proportion of features deactivated by the Dropout layer, preferably 0.1, that is, 10% of the input features are temporarily deactivated randomly during training. The weight decay coefficient λ is the weight penalty parameter in regularization, preferably 1e-4, which prevents overfitting caused by excessive weights by applying L2 penalty to the model weights.

[0112] (4) Number of training rounds: The maximum number of iterations or training rounds for model training is set to 100 rounds to balance training sufficiency and computational efficiency. An early stopping strategy is adopted (training stops if the validation set loss does not decrease for 10 consecutive rounds). The early stopping strategy is an adaptive termination rule for model training. Training stops when the validation set loss does not decrease for 10 consecutive rounds to avoid invalid iterations and overfitting.

[0113] (5) Batch size: The number of training samples used in a single parameter update, set to 64 to balance training efficiency and gradient estimation accuracy. Loss function: A hybrid loss function is used to train the prediction model, combining mean squared error (MSE) and mean absolute percentage error (MAPE), as shown in the following formula:

[0114]

[0115] in, The model training uses a hybrid loss function that combines mean squared error (MSE) and mean absolute percentage error (MAPE) to balance the prediction errors of different magnitudes of power meteorological elements (such as wind speed and icing thickness). The weighting coefficients of mean squared error (MSE) in the mixed loss function, preferably... =0.7, highlighting the penalizing effect of large errors on training. The weighting coefficient of the mean absolute percentage error (MAPE) in the mixed loss function, preferably... =0.3, adapting to error assessment of different magnitudes of factors, improving the consistency of prediction accuracy. MSE ( Mean squared error loss (MAPE) is calculated as the squared mean of the differences between the true and predicted values; it is sensitive to large errors. Mean absolute percentage error loss: Calculates the average percentage of the absolute error between the true value and the predicted value relative to the true value, adapting to error assessment of different magnitudes of factors; : The actual values ​​of meteorological elements related to power (such as measured wind speed, actual icing thickness, etc.); : The predicted values ​​of power meteorological elements output by the model.

[0116] S3.2.3: Training Process

[0117] 1. Initialize TAMS-RNNs model parameters ( (etc.) follows a random normal distribution. Model parameter initialization is the parameter assignment operation before model training, which... The learnable parameters are initialized to a random normal distribution to provide a reasonable starting point for training. : The input weight matrix corresponding to the k-th small hidden state in the multi-scale feature decoupling layer (MSFD), which is used to map the input features to the hidden state space of this scale. : The recursive matrix corresponding to the k-th small hidden state in the multi-scale feature decoupling layer (MSFD), used to model the temporal dependencies at this scale. The weight matrix corresponding to the input features in the Time-Aware Feature Modulation (TAFM) layer is used to generate scale importance weights. The weight matrix corresponding to the historical hidden states in the Time-Aware Feature Modulation (TAFM) layer is used to fuse historical context to generate scale weights.

[0118] 2. The input is a standardized multi-temporal-scale feature vector (containing 42-dimensional features, covering space-ground-network data). The input sequence length is adjusted according to the forecast lead time (Ultra-short-term T=144, short-term T=72, medium-term T=10). The training set data is input into the model in time series order. The multi-scale features are decoupled through the MSFD layer, and the scale weights are dynamically modulated by the TAFM layer. The standardized multi-temporal-scale feature vector is the preprocessed feature data input to the model. It is Z-score standardized to eliminate dimensional differences and contains 42-dimensional features (covering multi-source data from space, ground, and network). The input sequence length T is the time step of the input features: ultra-short-term forecast T=144 (24 hours × 6 minutes), short-term forecast T=72 (72 hours × 1 hour), and medium-term forecast T=10 (10 days × 1 day), adapting to different forecast lead time requirements.

[0119] 3. Calculate the mixed loss L between the predicted and actual values, and update the model parameters using the backpropagation algorithm. The backpropagation algorithm is the core algorithm for updating model parameters. By calculating the gradient of the loss function L with respect to each parameter, the parameter values ​​are adjusted along the gradient descent direction to minimize the prediction error.

[0120] 4. After each round of training, evaluate the model performance using the validation set (indicators are RMSE and MAPE), and record the optimal model parameters.

[0121] 5. The output is a customized forecast result for transmission lines, new energy power plants, and power distribution scenarios, including 12 key power meteorological elements: wind speed, wind direction, gust intensity, temperature, humidity, air pressure, precipitation intensity, snow depth, ice thickness, total irradiance, dust concentration, and thunderstorm probability. The output resolution reaches 100 meters (spatial) and minutes (time).

[0122] This invention, in step S3, constructs and trains a deep learning prediction model based on Time-Aware Multi-Scale Recurrent Neural Networks (TAMS-RNNs). Through feature decoupling and time-aware modulation mechanisms, it adaptively captures the multi-temporal and spatial scale dependencies of power meteorological data, achieving collaborative learning of features at different time and spatial scales and improving the prediction accuracy of meteorological elements in complex terrain areas. A multi-scale feature decoupling layer (MSFD) is constructed, decomposing the single hidden state of a traditional RNN into multiple independently updated small hidden states. Differentiated update frequencies are used to adapt to features at multiple time scales, such as minute, hour, day, and seasonal levels, solving the problem of inaccurate multi-scale modeling caused by feature entanglement in traditional models. A time-aware feature modulation layer (TAFM) is constructed, dynamically generating scale importance weights based on temporal context to achieve adaptive selection of the optimal scale under different forecast scenarios, matching differentiated operational needs such as ultra-short-term sudden weather events, short-term power grid dispatching, and medium-term maintenance plans. Finally, the trained prediction model is obtained and used for subsequent data fusion analysis and refined forecasting of various power meteorological elements.

[0123] S4. Input the feature data of the area to be predicted into the prediction model trained in step S3, and output the data fusion analysis results of multiple power meteorological elements in the future specified time period.

[0124] Based on the trained TAMS-RNNs model, forecasts of key power meteorological elements in multiple scenarios are achieved, such as... Figure 4 The image shown is a schematic diagram of the weather forecast visualization results according to an embodiment of the present invention:

[0125] 1. Ultra-short-term forecast (0–6 hours): Focuses on predicting rapidly evolving factors such as line icing, thunderstorms, and gusts to support real-time power grid dispatch. Specifically, in response to ultra-short-term forecast scenarios, the prediction model takes into account high-frequency feature sequences from the past 1 or 6 hours and outputs prediction results for the next 0–6 hours with a time resolution of minutes.

[0126] 2. Short-term forecast (1-3 days): Focuses on forecasting factors such as wind speed, irradiance, and temperature to support new energy power forecasting and grid operation planning. Specifically, in response to short-term forecast scenarios, the forecasting model takes into account the feature sequence of the past 72 hours and outputs forecast results for the next 1-3 days with an hourly time resolution.

[0127] 3. Medium-term forecast (3-10 days): Focuses on predicting the probability of extreme weather, precipitation, and average meteorological conditions to support equipment maintenance plans and risk assessments. Specifically, in response to the medium-term forecast scenario, the prediction model takes into account the daily average feature sequence of the past 10 days and outputs the prediction results for the next 3-10 days with a daily time resolution.

[0128] In summary, this invention employs a Time-Aware Multi-Scale Recurrent Neural Network (TAMS-RNNs) as its core innovation. It utilizes a Multi-Scale Feature Decoupling Layer (MSFD) to split the hidden state into four independent states with differentiated update frequencies, thus resolving the feature entanglement problem. Furthermore, it uses a Time-Aware Feature Modulation Layer (TAFM) to dynamically generate and modulate the weights of each state based on the temporal context, achieving precise adaptation to different forecast scenarios. This method combines multi-source data fusion from the air, ground, and network with multi-temporal and spatial scale feature construction, ultimately achieving high-resolution (100-meter spatial level, minute-level temporal level) refined forecasts for 12 key power meteorological elements, significantly improving the accuracy and operational adaptability of meteorological forecasts in complex terrain areas.

[0129] like Figure 5The figure shows a comparison of the accuracy of the prediction model of this invention with existing prediction models. The test experiment used the mean absolute percentage error (MAPE) as the evaluation index, comparing the convergence characteristics and final prediction accuracy of eight algorithms—including the basic recurrent neural network (RNN), long short-term memory network (LSTM), and variants fused with MSFD / TAFM modules—during 50 training iterations. The results show that the MAPE of all algorithms continuously decreases and gradually converges with the increase of iterations, verifying the learning ability of each model for the temporal characteristics of power meteorology. The MAPE of the basic RNN and LSTM models after convergence are 0.15 and 0.18, respectively, reflecting the limited fitting ability of a single network structure for multi-temporal and spatial scale power meteorological data. Algorithms fusing single modules (MSFD or TAFM) (such as RNN+MSFD, LSTM+TAFM) have a MAPE reduction of 10%~33% compared to the basic model, proving that the multi-scale feature separation mechanism of MSFD and the dynamic weight modulation strategy of TAFM can effectively improve the model's ability to capture complex meteorological characteristics in Northwest China. Meanwhile, algorithms that integrate MSFD and TAFM modules (LSTM+MSFD+TAFM, RNN+MSFD+TAFM) exhibit superior performance. Among them, the RNN+MSFD+TAFM algorithm proposed in this invention has a MAPE as low as 0.078 (<0.08) after convergence, which is further reduced compared to LSTM+MSFD+TAFM (0.09) and RNN+MSFD (0.12). This fully demonstrates the synergistic effect of MSFD and TAFM modules in the RNN architecture. MSFD decouples the hidden state into multi-scale sub-states to adapt to the evolution of meteorological elements at different time scales, while TAFM dynamically adjusts the weights of each scale based on the temporal context, accurately matching the differentiated needs of ultra-short-term, short-term, and medium-term power meteorological forecasts. Ultimately, the model achieves a lower error level and better convergence characteristics in the power meteorological element prediction task in Northwest China.

[0130] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0131] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for fusing and analyzing multi-temporal and spatial scale power meteorological data based on deep learning, characterized in that: Includes the following steps: S1. Collect and fuse multi-source heterogeneous data from the sky layer, ground layer, and network layer, and construct a unified power meteorological feature dataset after data preprocessing; S2. Based on the power meteorological feature dataset, construct multi-timescale features that adapt to the ultra-short-term, short-term and medium-term forecasting needs of power business, as well as multi-spatial-scale features covering the area, station and equipment scales. S3. Construct and train a prediction model based on a time-aware multi-scale recurrent neural network, wherein the prediction model is used to perform fusion analysis on the multi-time-scale features and the multi-spatial-scale features: decouple the hidden states formed by the recurrent unit in the process of processing time-series data into multiple parallel small hidden states for representing the evolution characteristics of different time scales; configure an update mechanism with independent and differentiated update frequencies for each small hidden state; dynamically determine the importance weight of each small hidden state to the current prediction based on the current time-series input information and historical context, and modulate the corresponding historical small hidden state information according to the importance weight; generate the final power meteorological element prediction result based on the modulated small hidden states. S4. Input the feature data of the area to be predicted into the prediction model trained in step S3, and output the prediction results of various power meteorological elements in the future specified time period.

2. The method according to claim 1, characterized in that, In step S1, the sky layer data includes meteorological satellite remote sensing data and weather radar data; the ground layer data includes meteorological station observation data distributed at different terrains and altitudes, new energy power plant monitoring data, transmission line micro-meteorological monitoring data, and special monitoring data for severe weather; the network layer data includes power grid operation status data, power equipment fault records, power operation safety data, and power grid geographic information data.

3. The method according to claim 1, characterized in that, In step S3, the prediction model includes an input layer, a multi-scale feature decoupling layer, a time-aware feature modulation layer, a fusion layer, and an output layer connected in sequence. The input layer is used to receive an input feature sequence consisting of multi-time-scale features and multi-spatial-scale features; The multi-scale feature decoupling layer is used to split the hidden state of the previous time step into K independent small hidden states, and control each small hidden state to be selectively updated according to its preset update frequency. The selective update includes keeping the state unchanged at non-update times and calculating the state based on the current input and the historical value corresponding to the small hidden state at update times. The time-aware feature modulation layer is used to calculate the dynamic weights corresponding to each small hidden state based on the input features of the current time step and the complete hidden state of the previous time step, and to use the dynamic weights to weight the corresponding historical small hidden state values ​​involved in the calculation when the multi-scale feature decoupling layer performs update calculation. The fusion layer is used to fuse multiple modulated small hidden states to form a unified global feature representation; The output layer is used to map the global feature representation into predicted values ​​of various power meteorological elements.

4. The method according to claim 3, characterized in that, The specific operations of the multi-scale feature decoupling layer include: Hidden state at time step t Split into K small hidden states k=1,2,…,K, each small hidden state Corresponding to an update scale ; For the k-th small hidden state at time step t Its update rules are as follows: If the current time step t is not an update scale If the value is an integer multiple of the previous time step, then the current state copies the state from the previous time step. Otherwise, update based on the input feature vector at time t and the state at the previous time step, i.e. ; in, Let be the input feature vector at time t. For the input weight matrix, For a recursive matrix, Here, tanh is the bias term, and tanh is the hyperbolic tangent activation function. This represents the k-th small hidden state at the (t-1)-th time step.

5. The method according to claim 4, characterized in that, The number of small hidden states is K=4, and the update scale set S is... ; It is used to capture time scale dependencies at the minute / hour, hour / day, day / week, and week / seasonal levels.

6. The method according to claim 5, characterized in that, The time-aware feature modulation layer calculates the dynamic weights through a weight generation unit. The weight generation unit receives the input feature vector of the current time step and the complete hidden state vector of the previous time step, performs linear transformation and normalization, and outputs a set of weight values ​​that sum to 1. Each weight value corresponds to a small hidden state.

7. The method according to claim 1, characterized in that, In step S3, the prediction model is trained using a hybrid loss function L, which is a weighted sum of the mean squared error (MSE) and the mean absolute percentage error (MAPE). ； in, For the true value, For predicted values, and These are the weighting coefficients.

8. The method according to claim 1, characterized in that, In step S4, the power meteorological elements include at least two of the following: wind speed, wind direction, gust intensity, temperature, humidity, air pressure, precipitation intensity, snow depth, ice thickness, total irradiance, dust concentration, and thunderstorm probability.

9. The method according to claim 8, characterized in that, Step S4 specifically includes: In response to ultra-short-term forecast scenarios, the prediction model takes into input a high-frequency feature sequence of the past 1 hour or 6 hours and outputs a prediction result for the next 0-6 hours with a time resolution of minutes. In response to short-term forecast scenarios, the prediction model takes a feature sequence of the past 72 hours as input and outputs prediction results for the next 1-3 days with a time resolution of hours. In response to medium-term forecast scenarios, the prediction model takes into account the daily average feature sequence of the past 10 days and outputs the prediction results for the next 3-10 days with a time resolution of daily.

10. The method according to claim 8, characterized in that, In step S2, the multi-spatial-scale features include: regional-scale features covering a province or river basin, station-scale features covering new energy power stations, and equipment-scale features covering transmission lines or substations; in step S1, data preprocessing includes spatial interpolation, which is used to downscale regional-scale meteorological data to the equipment scale to match the spatial distribution of power grid equipment, wherein the spatial resolution of the equipment scale is 100 meters.

Images