A multi-source meteorological and power data fusion processing method

Abstract

The application provides a multi-source meteorological and power data fusion processing method.The multi-source meteorological and power data fusion processing method comprises the following steps: S1, multi-source heterogeneous meteorological data acquisition and space-time reference unification; S2, physical consistency constraint pre-screening; S3, weather type adaptive skill matrix construction; S4, dynamic confidence weight fusion; S5, terrain perception local residual correction step; S6, uncertainty quantification transmission and source level health degradation step; and S7, meteorological-power mapping.The multi-source meteorological and power data fusion processing method has the advantages that the accuracy and reliability of new energy power prediction can be significantly improved through weather type adaptive dynamic weight fusion, physical consistency constraint pre-screening, terrain perception local correction, uncertainty quantification transmission and source level health degradation.

Description

Technical Field

[0001] This invention relates to the field of data fusion technology, and in particular to a method for fusion processing of multi-source meteorological and power data. Background Technology

[0002] With the deepening of the "dual-carbon" goals and the accelerated construction of new power systems, the proportion of renewable energy generation is increasing daily. However, wind and solar power output is greatly affected by weather conditions, and their volatility and intermittency pose a severe challenge to the safe and stable operation of the power system. Accurate renewable energy power forecasting is a key support for achieving grid dispatching, electricity market trading, and lean management of power plants.

[0003] Currently, new energy power forecasting typically relies on numerical weather prediction products (such as ECMWF, CMA-WSP, etc.) as meteorological input, and then uses statistical or machine learning models to establish a mapping relationship between meteorological elements and power generation. However, existing technologies have the following significant drawbacks:

[0004] The current multi-source meteorological data fusion methods are crude: In practical applications, multiple numerical weather prediction models, satellite inversion products, and ground observation data often coexist. However, existing fusion methods mostly use fixed weights or simple arithmetic averages, failing to consider the performance differences of different data sources under different weather types and terrain conditions. For example, some models have high accuracy under clear weather conditions, but their errors increase sharply under typhoons or severe convective weather. Fixed-weight fusion will bring low-quality data into the final result, leading to a decrease in prediction accuracy.

[0005] Lack of physical consistency verification: Multi-source data are not rigorously screened for physical constraints before being fused, and may contain outliers that violate basic physical laws (such as solar radiation greater than zero at night, or serious discrepancies between wind speed and air pressure). These physical outliers not only contaminate the fusion results, but also mislead the training of subsequent machine learning models, causing them to learn incorrect mapping relationships.

[0006] Insufficient correction capability for meteorological elements in complex terrain: In complex terrain areas such as karst landforms, mountains, and hills, the coarse resolution of numerical weather prediction models is insufficient to characterize the effects of local topography, such as convergence, airflow around the terrain, valley winds, and shading. Existing technologies either ignore the influence of topography or use simple empirical formulas for correction, resulting in limited accuracy and poor generalization ability.

[0007] Lack of forecast uncertainty: Most power forecasting systems only output a single point forecast value without providing a confidence interval for that value. Dispatchers cannot determine the reliability of the current forecast, making them prone to making incorrect decisions during extreme weather or when data is abnormal.

[0008] Weak response mechanisms for data source failures: When a data source experiences a quality degradation due to communication interruption, instrument failure, or calibration drift, the existing system typically passively receives inferior data or directly reports an error and interrupts the process, lacking an intelligent fault-tolerant mechanism for automatic detection, degradation, and switching to backup links.

[0009] Therefore, it is necessary to provide a new method for fusing multi-source meteorological and power data to solve the above-mentioned technical problems. Summary of the Invention

[0010] The technical problem solved by this invention is to provide a multi-source meteorological and power data fusion processing method that can significantly improve the accuracy and reliability of new energy power prediction by using weather-type adaptive dynamic weight fusion, physical consistency constraint pre-screening, terrain perception local correction, uncertainty quantification and transmission, and source-level health degradation.

[0011] To solve the above-mentioned technical problems, the multi-source meteorological and power data fusion processing method provided by the present invention includes the following steps:

[0012] S1. Acquire meteorological data from multiple sources, including numerical weather prediction products, meteorological satellite remote sensing data, meteorological radar echo data, and measured data from ground automatic weather stations. Interpolate the meteorological data from these multiple sources onto the same spatiotemporal reference grid to generate a spatiotemporally aligned multi-source meteorological dataset.

[0013] S2. Perform physical layer hard constraint verification on meteorological elements of each grid point or station of each data source in the spatiotemporally aligned multi-source meteorological dataset. The physical layer hard constraints include: dew point temperature not exceeding dry bulb temperature, surface solar radiation irradiance being non-negative and not exceeding the theoretical upper limit of the top of the atmosphere, and near-surface wind speed and air pressure satisfying the rationality criteria of static adaptability. Data points that violate any physical layer hard constraints are marked as physical outliers and isolated to generate a multi-source meteorological dataset after physical pre-screening.

[0014] S3. Extract the weather circulation feature vectors corresponding to the multi-source meteorological datasets after physical pre-screening, classify the weather type for the current period based on the weather circulation feature vectors, and obtain the current weather type category; for each weather type category, construct a time-varying skill scoring matrix for each data source relative to the weather type based on the historical observation data of the same period and the prediction deviation statistics of each data source under the weather type, and the time-varying skill scoring matrix quantifies the relative credibility of each data source on each meteorological element.

[0015] S4. Index the corresponding time-varying skill score matrix according to the current weather type category, and assign dynamic confidence weights to each data source. The dynamic confidence weights are positively correlated with the skill score of the data source under the current weather type and negatively correlated with the data dispersion of the data source within the current sliding time window. Based on the dynamic confidence weights, perform weighted average fusion of the same meteorological element in each data source in the multi-source meteorological dataset after physical pre-screening to generate a fused meteorological field.

[0016] S5. Obtain the static terrain data of the station location. The static terrain data includes elevation, slope, aspect and surface roughness parameters. Using the static terrain data as input, output the local deviation correction field of terrain perception through the pre-trained terrain-deviation mapping model. Use the local deviation correction field to perform grid-by-grid residual correction on the fused meteorological field to obtain a high-precision meteorological field after terrain correction.

[0017] S6. Calculate the deviation divergence of each data source relative to the fusion result during the weighted average fusion process, estimate the prediction uncertainty of each grid point in the high-precision meteorological field based on the deviation divergence, and output meteorological element prediction products with confidence intervals; at the same time, determine the health level of each data source according to the duration of the deviation divergence exceeding the preset threshold, and when the health level of any data source drops below the degradation threshold, remove the data source from the fusion weight and trigger the backup link to achieve degradation prediction without interrupting the service.

[0018] S7. The meteorological element prediction product with confidence interval is spatiotemporally matched with the historical power data, equipment parameters and real-time operating status data of the new energy power station. By embedding a machine learning prediction model with power conversion physical constraints, the multi-timescale power prediction value and the corresponding power prediction confidence interval of each power station are output.

[0019] Preferably, in step S3, the weather circulation feature vector is extracted from the statistics of geopotential height field, sea level pressure field and near-surface temperature gradient of numerical weather forecast products. The weather type classification uses an unsupervised clustering algorithm to divide the circulation feature vectors of all historical time periods into several typical weather type categories. The typical weather type categories distinguish between large-scale stable type, frontal passage type, strong convection type and typhoon or tropical cyclone influence type.

[0020] Preferably, the specific method for constructing the time-varying skill score matrix in S3 is as follows: for any data source, any meteorological element, and any weather type, the skill score is defined as the inverse normalized value of the root mean square error of the cross-validation of the element under that weather type; the dynamic confidence weight is determined by normalizing the product of the skill score and an exponential decay factor based on the data dispersion within the current sliding window, so that the data source with greater dispersion receives a smaller weight.

[0021] Preferably, the physical layer hard constraint verification in S2 further includes: for photovoltaic power plants, verifying that the total solar radiation on the ground surface during the daytime meets the radiation component consistency relationship with the direct radiation component and the scattered radiation component, that is, the total radiation is not less than the direct radiation component multiplied by the zenith angle cosine and the scattered radiation component; for wind farms, verifying whether the hub height wind speed is compatible with the logarithmic law or power law profile of the near-surface wind shear formula, and values ​​exceeding the physical allowable range are marked as physical anomalies.

[0022] Preferably, the terrain-bias mapping model in S5 is constructed as follows: the static terrain data is encoded into a terrain feature vector including elevation, slope modulus, aspect, surface roughness, and land cover category; using the residual between ground station observations and the fused background field over a historical period as the supervision target, a regression model or lightweight neural network is trained with the terrain feature vector and the local gradient of the fused meteorological field as input, and the output is the terrain-aware local bias correction field. The final correction result is the fused background field plus the local bias correction field.

[0023] Preferably, in step S6, the specific method for estimating the prediction uncertainty is as follows: the fusion variance of each data source after dynamic weight fusion is approximately estimated as the sum of the product of the square of the weights of each data source and its own variance, plus half of the sum of the product of the weights of each data source and the product of the residual covariance; based on this, the prediction interval of the meteorological element is output, and the prediction interval is transmitted to step S7 along with the predicted value of the meteorological element, so that the power prediction confidence interval is obtained by the joint propagation of meteorological uncertainty and power conversion model error.

[0024] Preferably, in step S6, the downgrading of the health level of the data source includes the following steps: setting a health score for each data source, initialized to the highest value; within each update cycle, when the proportion of data sources marked as physical anomalies exceeds a first threshold, or when the duration for which their deviation divergence continuously exceeds a second threshold exceeds a set duration, the health score is reduced according to a decay function; when the health score is lower than the health threshold, the following operations are performed: first, resetting the dynamic confidence weight of the data source to zero, and renormalizing the weights of the remaining data sources; second, activating the backup link corresponding to the data source—the backup link includes one or more combinations of: switching to a similar backup source, spatial interpolation filling of neighboring stations, or calling an integrated alternative scheme of a set of historical days with similar weather patterns; third, generating a data source downgrading event log for display by the central control center monitoring module.

[0025] Preferably, the S7 incorporates a machine learning prediction model with physical constraints on power conversion. Its training objective function includes a physical regularization term, which comprises: a penalty term for forcing non-negativity of the predicted power, a penalty term for constraining the predicted power from exceeding the theoretical capacity limit, and a monotonicity regularization term constraining the predicted power to have a physically monotonically non-decreasing characteristic with respect to wind speed or irradiance under the same unit conditions. In addition to the wind speed, irradiance, temperature, and air pressure elements from the high-precision meteorological field after terrain correction, the model input also includes a variance estimate of the prediction uncertainty as an uncertainty perception channel, enabling the model to automatically output a wider power prediction interval in the non-confidence region.

[0026] Preferably, in step S3, the time-varying skill scoring matrix is ​​updated online using a rolling time window: whenever new ground station measured data is acquired, the root mean square error of cross-validation and the standard deviation of element observations of the data source are updated under the corresponding weather type category, thereby realizing the continuous adaptive evolution of the dynamic confidence weights, rather than relying on statically set fixed fusion weights offline.

[0027] Preferably, in step S1, for numerical forecast grid points with different spatial resolutions and irregularly distributed ground station data, a regrinding interpolation method that maintains the conservation of physical quantities is adopted—bilinear interpolation is performed on the wind speed vector component on the horizontal plane before synthesizing the modulus, area-weighted average interpolation is performed on scalar quantities such as radiation flux, and mass-weighted interpolation is performed on state quantities such as temperature, pressure and humidity, so as to avoid introducing false energy or flux imbalances due to improper interpolation methods.

[0028] Compared with related technologies, the multi-source meteorological and power data fusion processing method provided by this invention has the following beneficial effects:

[0029] This invention provides a method for fusing multi-source meteorological and power data. Through a weather-adaptive skill matrix and dynamic confidence weights, it enables differentiated fusion of multi-source meteorological data under various weather conditions, effectively avoiding the accuracy loss caused by fixed-weight fusion and improving fusion quality under complex weather conditions. Physical consistency constraint pre-screening removes non-physical anomalies at the data entry point, preventing inferior data from entering subsequent stages and enhancing the robustness of the entire prediction chain. A terrain-aware local residual correction model captures terrain-induced local meteorological deviations, significantly improving the consistency between the corrected meteorological field and measured data in complex terrain areas. It outputs meteorological and power prediction products with confidence intervals, providing a quantitative basis for scheduling decisions and helping to optimize strategies in power market transactions. A source-level health degradation mechanism enables real-time monitoring and automatic fault tolerance of data source quality. When a data source fails or deteriorates, the system can seamlessly switch to a backup link, ensuring uninterrupted prediction services, making it particularly suitable for unattended control center operation and maintenance scenarios. Attached Figure Description

[0030] Figure 1 The flowchart shows the multi-source meteorological and power data fusion processing method provided by the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] Please refer to the following: Figure 1 ,in, Figure 1 A flowchart illustrating the multi-source meteorological and power data fusion processing method provided by this invention. The multi-source meteorological and power data fusion processing method includes the following steps:

[0033] S1. Acquire meteorological data from multiple sources, including at least numerical weather prediction products, meteorological satellite remote sensing data, meteorological radar echo data, and measured data from ground automatic weather stations. Interpolate the meteorological data from these multiple sources onto the same spatiotemporal reference grid to generate a spatiotemporally aligned multi-source meteorological dataset.

[0034] Among them, for numerical forecast grid points with different spatial resolutions and irregularly distributed ground station data, regrinding interpolation that maintains the conservation of physical quantities is adopted—bilinear interpolation is performed on the wind speed vector components on the horizontal plane and then the modulus is synthesized; area-weighted average interpolation is performed on scalar quantities such as radiation flux; and mass-weighted interpolation is performed on state quantities such as temperature, pressure and humidity, so as to avoid introducing false energy or flux imbalances due to improper interpolation methods.

[0035] S2. Perform physical layer hard constraint verification on meteorological elements of each grid point or station of each data source in the spatiotemporally aligned multi-source meteorological dataset. The physical layer hard constraints include at least the following: dew point temperature does not exceed dry bulb temperature, surface solar irradiance is non-negative and does not exceed the theoretical upper limit of the top of the atmosphere, and near-surface wind speed and air pressure satisfy the rationality criteria for static adaptability. In addition, for photovoltaic stations, verify that the total surface solar irradiance GHI during the daytime is consistent with the direct radiation DNI and the diffuse radiation DHI. The radiative component consistency relationship is given by GHI, where GHI is the total irradiance of the horizontal plane in watts per square meter; DNI is the direct normal irradiance in watts per square meter; and DHI is the diffuse irradiance of the horizontal plane in watts per square meter. The solar zenith angle is expressed in radians. For the wind speed at the hub height of the wind farm, we verify whether it is compatible with the logarithmic law or power law profile of the near-surface wind shear formula. Values ​​that exceed the physical allowable range are marked as physical outliers. Data points that violate any physical layer hard constraints are marked as physical outliers and isolated. This generates a multi-source meteorological dataset that has been pre-screened by physical analysis.

[0036] S3. Extract the weather circulation feature vectors corresponding to the multi-source meteorological datasets after physical pre-screening. The weather circulation feature vectors are extracted from at least the statistics of geopotential height field, sea level pressure field, and near-surface temperature gradient of numerical weather prediction products. Based on the weather circulation feature vectors, classify the weather type for the current period. The weather type classification adopts the K-means unsupervised clustering algorithm, and determines the number of clusters K by the elbow rule. Divide the circulation feature vectors of all historical time periods into several typical weather type categories. The typical weather type categories should at least distinguish between large-scale stable type, frontal passage type, strong convection type, and typhoon or tropical cyclone impact type to obtain the current weather type category. For each weather type category, construct a time-varying skill scoring matrix for each data source relative to the weather type based on the prediction deviation statistics of historical observation data and each data source under the weather type. The time-varying skill scoring matrix quantifies the relative reliability of each data source on each meteorological element.

[0037] Among them, the skill score for data source i, meteorological element v, and weather type k is defined as follows:

[0038]

[0039] in, This represents the skill score of data source i for meteorological element v under weather type k. The value ranges from 0 to 1. The larger the value, the more reliable the data source is under that weather type. This represents the root mean square error of the cross-validation of the data source for element v under weather type k, expressed in physical units of meteorological elements. This represents the observed standard deviation of meteorological element v under weather type k, with units in the physical units of the meteorological element.

[0040] Furthermore, the time-varying skill scoring matrix is ​​updated online using a rolling time window: whenever new ground station measurement data is acquired, the data source is updated under the corresponding weather type category. and This enables the continuous adaptive evolution of the dynamic confidence weights, rather than relying on statically set fixed fusion weights offline.

[0041] S4. Index the corresponding time-varying skill score matrix according to the current weather type category, and assign dynamic confidence weights to each data source. The dynamic confidence weights are positively correlated with the skill score of the data source under the current weather type and negatively correlated with the data dispersion of the data source within the current sliding time window. Based on the dynamic confidence weights, perform weighted average fusion of the same meteorological element in each data source in the multi-source meteorological dataset after physical pre-screening to generate a fused meteorological field.

[0042] Wherein, the dynamic confidence weight It is given by the following formula:

[0043]

[0044] in, Let be the fusion weight of data source i at the current moment, dimensionless, with the sum of all weights equal to 1; similarly, For data source i, the skill rating of element v under weather type k; This is the discrete penalty coefficient, which is dimensionless and takes a positive value. It is used to control the intensity of the penalty imposed on the weights by the discreteness. This measures the dispersion of elements in data source i within the current sliding time window, in physical units of meteorological elements, representing the degree of fluctuation of the recent forecast values ​​of this data source relative to its historical average; the summation sign in the denominator... This represents the summation over all data sources j participating in the fusion;

[0045] S5. Obtain the static terrain data of the station location. The static terrain data includes at least elevation, slope, aspect and surface roughness parameters. Using the static terrain data as input, output the local deviation correction field of terrain perception through the pre-trained terrain-deviation mapping model. Use the local deviation correction field to perform grid-by-grid residual correction on the fused meteorological field to obtain a high-precision meteorological field after terrain correction.

[0046] The terrain-bias mapping model employs a three-layer fully connected neural network structure. The input layer has 5 nodes equal to the terrain feature vector dimension, while the hidden layers have 64 and 32 nodes respectively. The ReLU activation function is used, and the output layer has 1 node, outputting the terrain-aware local bias correction value. Model training utilizes the Adam optimizer with a learning rate of 0.001, a batch size of 128, an early stopping epoch of 10, and a maximum training epoch of 200. Training data is derived from the residuals between ground station observations and the fused background field over the past three years, divided into training and validation sets in an 8:2 ratio.

[0047] Terrain feature vector The encoding is:

[0048]

[0049] Where (x,y) are the latitude and longitude coordinates of the grid point; h(x,y) is the elevation of the grid point in meters; ∇h(x,y) is the slope modulus, which is dimensionless; aspect(x,y) is the slope aspect in degrees. LC(x,y) represents the surface roughness length in meters; LC(x,y) represents the land cover category, a categorical variable; and the residual between ground station observations and the fused background field over a historical period is used.

[0050] For the purpose of supervision, among which Let be the observation value of the ground station at time t for grid point (x,y). The values ​​of the fused background field output by S4 at the same grid point and time are... The units are the same as those for meteorological elements; a regression model or lightweight neural network is trained with T and the local gradient of the fused meteorological field as inputs, and the output is the local bias correction field of the terrain perception. , The units are the same as those for meteorological elements; the final correction result is:

[0051]

[0052] in, This represents the value of the high-precision meteorological field after terrain correction at grid point (x,y). The value of the fused background field output by S4 at the grid point (x,y) is given. This is the local bias correction value for terrain perception.

[0053] S6. Calculate the deviation divergence of each data source relative to the fusion result during the weighted average fusion process, estimate the prediction uncertainty of each grid point in the high-precision meteorological field based on the deviation divergence, and output meteorological element prediction products with confidence intervals; at the same time, determine the health level of each data source according to the duration of the deviation divergence exceeding the preset threshold, and when the health level of any data source drops below the degradation threshold, remove the data source from the fusion weight and trigger the backup link to achieve degradation prediction without interrupting the service.

[0054] The specific method for estimating the prediction uncertainty is as follows: the fusion variance of each data source after dynamic weight fusion is approximately estimated as follows:

[0055]

[0056] in, This is the variance estimate after fusion at grid point (x,y), in squares of meteorological element units; Let be the fusion weight of data source i at grid point (x,y), which is dimensionless; The variance estimate of the elements at grid point (x,y) of data source i is calculated from the variance of the historical forecast error of the data source, and the unit is the square of the meteorological element unit; Let $\frac{i}{j}$ be the residual covariance between data source $i$ and data source $j$ at grid point $(x,y)$, in squared units of meteorological elements; the first summation symbol is $\frac{i}{j}$. Summing over all data sources i, the second double summation Summate all distinct data source pairs (i,j).

[0057] Based on this, the forecast range for meteorological elements is:

[0058]

[0059] in, These are the meteorological element values ​​after topographic correction; To integrate standard deviation, i.e. ; The quantile of the standard normal distribution is γ, and the significance level is γ (for example, 0.05 corresponds to a 95% confidence interval, z0.975≈1.96); the prediction interval is then passed into S7 along with the predicted values ​​of meteorological elements, so that the power prediction confidence interval is obtained by the joint propagation of meteorological uncertainty and power conversion model error;

[0060] The source-level health downgrade step specifically includes: setting a health score for each data source. Initialized to 1; within each update cycle, when the proportion of data source i marked as physical outliers exceeds a first threshold. or its deviation divergence Exceeding the second threshold consecutively The duration exceeds At that time, Decrease according to the decay function; when At that time, perform the following operations: First, adjust the dynamic confidence weight of the data source. First, set the weights of the remaining data sources to zero and renormalize them. Second, activate the backup link corresponding to the data source—the backup link includes one or more combinations of: switching to a similar backup source, spatial interpolation filling from neighboring stations, or calling an integrated alternative set of historical days with similar weather patterns. Third, generate a data source degradation event log for display by the central control center's monitoring module. The health score for data source i is dimensionless and ranges from 0 to 1, where 1 represents perfect health. This is a physical anomaly threshold, dimensionless, for example, 5%. This is the dispersion threshold, in units of 1 and 2. same; This is a duration threshold, in hours or minutes; Set a threshold for downgrading health status, dimensionless, for example, 0.3;

[0061] S7. The meteorological element prediction product with confidence interval is spatiotemporally matched with the historical power data, equipment parameters and real-time operating status data of the new energy power station. By embedding a machine learning prediction model with power conversion physical constraints, the multi-timescale power prediction value and the corresponding power prediction confidence interval of each power station are output.

[0062] The machine learning prediction model embedded with power conversion physical constraints uses a gradient boosting regression tree (XGBoost) as the base model, with a maximum tree depth of 6, a learning rate of 0.1, a subsampling ratio of 0.8, a column sampling ratio of 0.8, and 500 weak learners. The training objective function includes a physical regularization term, specifically in the form of:

[0063]

[0064] Where L is the total training loss function, which is dimensionless; To predict power Compared with actual power The mean square error between them is expressed as the square of the power units; , , These are the weight coefficients for each physical regularization term, dimensionless, and set to 0.1, 0.1, and 0.05 respectively; To penalize the expected square of the negative portion of the predicted power, the predicted power is forced to be non-negative; To predict power exceeding installed capacity A partial expected square penalty is applied to constrain the predicted power from exceeding the theoretical capacity limit. The monotonicity regularization term constrains the predicted power to have a physically monotonically non-decreasing characteristic with respect to wind speed or irradiance under the same unit conditions. In addition to the wind speed, irradiance, temperature, and air pressure elements from the high-precision meteorological field after terrain correction, the model input also includes the variance estimate of the prediction uncertainty. As an uncertainty perception channel, the model automatically outputs a wider power prediction interval in the non-confidence region. The training data uses the station operation data of the most recent three years, and time series cross-validation is performed by monthly sliding windows to ensure the model's generalization ability.

[0065] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for fusing and processing multi-source meteorological and power data, characterized in that, Includes the following steps: S1. Acquire meteorological data from multiple sources, including numerical weather prediction products, meteorological satellite remote sensing data, meteorological radar echo data, and measured data from ground automatic weather stations. Interpolate the meteorological data from these multiple sources onto the same spatiotemporal reference grid to generate a spatiotemporally aligned multi-source meteorological dataset. S2. Perform physical layer hard constraint verification on meteorological elements of each grid point or station of each data source in the spatiotemporally aligned multi-source meteorological dataset. The physical layer hard constraints include: dew point temperature not exceeding dry bulb temperature, surface solar radiation irradiance being non-negative and not exceeding the theoretical upper limit of the top of the atmosphere, and near-surface wind speed and air pressure satisfying the rationality criteria of static adaptability. Data points that violate any physical layer hard constraints are marked as physical outliers and isolated to generate a multi-source meteorological dataset after physical pre-screening. S3. Extract the weather circulation feature vectors corresponding to the multi-source meteorological datasets after physical pre-screening, classify the weather type for the current period based on the weather circulation feature vectors, and obtain the current weather type category; for each weather type category, construct a time-varying skill scoring matrix for each data source relative to the weather type based on the historical observation data of the same period and the prediction deviation statistics of each data source under the weather type, and the time-varying skill scoring matrix quantifies the relative credibility of each data source on each meteorological element. S4. Index the corresponding time-varying skill score matrix according to the current weather type category, and assign dynamic confidence weights to each data source. The dynamic confidence weights are positively correlated with the skill score of the data source under the current weather type and negatively correlated with the data dispersion of the data source within the current sliding time window. Based on the dynamic confidence weights, perform weighted average fusion of the same meteorological element in each data source in the multi-source meteorological dataset after physical pre-screening to generate a fused meteorological field. S5. Obtain the static terrain data of the station location. The static terrain data includes elevation, slope, aspect and surface roughness parameters. Using the static terrain data as input, output the local deviation correction field of terrain perception through the pre-trained terrain-deviation mapping model. Use the local deviation correction field to perform grid-by-grid residual correction on the fused meteorological field to obtain a high-precision meteorological field after terrain correction. S6. Calculate the deviation divergence of each data source relative to the fusion result during the weighted average fusion process, estimate the prediction uncertainty of each grid point in the high-precision meteorological field based on the deviation divergence, and output meteorological element prediction products with confidence intervals; at the same time, determine the health level of each data source according to the duration of the deviation divergence exceeding the preset threshold, and when the health level of any data source drops below the degradation threshold, remove the data source from the fusion weight and trigger the backup link to achieve degradation prediction without interrupting the service. S7. The meteorological element prediction product with confidence interval is spatiotemporally matched with the historical power data, equipment parameters and real-time operating status data of the new energy power station. By embedding a machine learning prediction model with power conversion physical constraints, the multi-timescale power prediction value and the corresponding power prediction confidence interval of each power station are output.

2. The multi-source meteorological and power data fusion processing method according to claim 1, characterized in that, In S3, the weather circulation feature vector is extracted from the statistics of geopotential height field, sea level pressure field and near-surface temperature gradient of numerical weather forecast products. The weather type classification uses an unsupervised clustering algorithm to divide the circulation feature vectors of all historical times into several typical weather type categories. The typical weather type categories distinguish between large-scale stable type, frontal passage type, strong convection type and typhoon or tropical cyclone influence type.

3. The multi-source meteorological and power data fusion processing method according to claim 2, characterized in that, The specific method for constructing the time-varying skill score matrix in S3 is as follows: for any data source, any meteorological element, and any weather type, the skill score is defined as the inverse normalized value of the root mean square error of the cross-validation of the element under that weather type; the dynamic confidence weight is determined by normalizing the product of the skill score and an exponential decay factor based on the data dispersion within the current sliding window, so that the data source with greater dispersion receives a smaller weight.

4. The multi-source meteorological and power data fusion processing method according to claim 1, characterized in that, The physical layer hard constraint verification in S2 also includes: for photovoltaic power plants, verifying that the total solar radiation on the ground surface during the daytime meets the radiation component consistency relationship with the direct radiation component and the scattered radiation component, that is, the total radiation is not less than the direct radiation component multiplied by the zenith angle cosine and the scattered radiation component; for wind farms, verifying whether the hub height wind speed is compatible with the logarithmic law or power law profile of the near-surface wind shear formula, and values ​​exceeding the physical allowable range are marked as physical anomalies.

5. The multi-source meteorological and power data fusion processing method according to claim 1, characterized in that, The terrain-bias mapping model in S5 is constructed as follows: static terrain data is encoded into terrain feature vectors including elevation, slope modulus, aspect, surface roughness, and land cover type; the residual between ground station observations and the fused background field over a historical period is used as the supervision target to train a regression model or lightweight neural network with the terrain feature vectors and the local gradient of the fused meteorological field as inputs, and outputs the terrain-aware local bias correction field. The final correction result is the fused background field plus the local bias correction field.

6. The multi-source meteorological and power data fusion processing method according to claim 1, characterized in that, In S6, the specific method for estimating the prediction uncertainty is as follows: the fusion variance of each data source after dynamic weight fusion is approximately estimated as the sum of the product of the square of the weight of each data source and its own variance, plus half of the sum of the product of the weights of each data source and the product of the residual covariance; based on this, the prediction interval of meteorological elements is output, and the prediction interval is transmitted to S7 along with the predicted value of meteorological elements, so that the power prediction confidence interval is obtained by the joint propagation of meteorological uncertainty and power conversion model error.

7. The multi-source meteorological and power data fusion processing method according to claim 1, characterized in that, In S6, the downgrading of the health level of a data source includes the following steps: setting a health score for each data source, initialized to the highest value; within each update cycle, when the proportion of data sources marked as physical anomalies exceeds a first threshold, or when the duration for which their deviation divergence continuously exceeds a second threshold exceeds a set duration, the health score is reduced according to a decay function; when the health score is lower than the health threshold, the following operations are performed: first, resetting the dynamic confidence weight of the data source to zero, and renormalizing the weights of the remaining data sources; second, activating the backup link corresponding to the data source—the backup link includes one or more combinations of: switching to a similar backup source, spatial interpolation filling of neighboring stations, or calling an integrated alternative scheme of historical weather-type similar day sets; third, generating a data source downgrading event log for display by the central control center monitoring module.

8. The multi-source meteorological and power data fusion processing method according to claim 1, characterized in that, The S7 incorporates a machine learning prediction model with physical constraints on power conversion. Its training objective function includes a physical regularization term, which includes: a penalty term to force the predicted power to be non-negative, a penalty term to constrain the predicted power from exceeding the theoretical capacity limit, and a monotonicity regularization term to constrain the predicted power to have a physically monotonically non-decreasing characteristic with respect to wind speed or irradiance under the same unit conditions. In addition to the wind speed, irradiance, temperature, and air pressure elements of the high-precision meteorological field after terrain correction, the model input also includes the variance estimate of the prediction uncertainty as an uncertainty perception channel, enabling the model to automatically output a wider power prediction interval in the non-confidence region.

9. The multi-source meteorological and power data fusion processing method according to claim 2, characterized in that, In S3, the time-varying skill scoring matrix is ​​updated online in a rolling time window manner: whenever new ground station measured data is obtained, the root mean square error of cross-validation and the standard deviation of element observations of the data source are updated under the corresponding weather type category, thereby realizing the continuous adaptive evolution of the dynamic confidence weight, rather than relying on the fixed fusion weight set statically offline.

10. The multi-source meteorological and power data fusion processing method according to claim 1, characterized in that, In S1, for numerical forecast grid points with different spatial resolutions and irregularly distributed ground station data, a regrinding interpolation method that maintains the conservation of physical quantities is adopted—bilinear interpolation is performed on the wind speed vector component on the horizontal plane before synthesizing the modulus, area-weighted average interpolation is performed on scalar quantities such as radiation flux, and mass-weighted interpolation is performed on state quantities such as temperature, pressure and humidity, so as to avoid introducing false energy or flux imbalances due to improper interpolation methods.

Images