Training method of power prediction model of new energy station and power prediction model

Abstract

The application discloses a training method of a power prediction model of a new energy station. The method comprises the following steps: performing data preprocessing on actual power and measured meteorological data corresponding to the new energy station to determine a power-meteorological alignment data set. The power-meteorological alignment data set is subjected to feature supplementing processing to determine a target power-meteorological alignment data set. A target training data is constructed according to a preset time window and the target power-meteorological alignment data set, wherein the preset time window comprises a history window and a prediction window. The preset model is trained according to the target training data to obtain the power prediction model. In this way, the target power-meteorological alignment data set is determined by performing feature supplementing processing on the power-meteorological alignment data set, and the supplemented features can help the preset model to learn the time sequence relationship that cannot be reflected by only relying on power or meteorological data, thereby improving the perception ability of the model to complex influencing factors in meteorological data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of new energy power prediction, and particularly relates to a training method of a power prediction model of a new energy station, the power prediction model, an electronic device and a computer readable storage medium. BACKGROUND

[0002] In the related art, the power prediction model is often used to predict the power generation of the new energy station under different meteorological data. However, the existing power prediction model often depends on the actual power and the basic meteorological data, and it is difficult to capture the rules between the new energy power generation and the meteorological data, so the power prediction accuracy is limited, and it is difficult to meet the actual demand of users for high-precision prediction. SUMMARY

[0003] The present application provides a training method of a power prediction model of a new energy station, the power prediction model, an electronic device and a computer readable storage medium.

[0004] The present application provides a training method of a power prediction model of a new energy station, the power prediction model, an electronic device and a computer readable storage medium.

[0005] The actual power and the measured meteorological data corresponding to the new energy station are preprocessed to determine a power-meteorological alignment data set.

[0006] The power-meteorological alignment data set is subjected to feature supplement processing to determine a target power-meteorological alignment data set.

[0007] According to the target training data, a preset model is trained to obtain the power prediction model.

[0008] According to the target training data, a preset model is trained to obtain the power prediction model.

[0009] Thus, data preprocessing is performed on the actual power and measured meteorological data corresponding to the new energy stations to determine the power-meteorological aligned dataset. Next, feature supplementation processing is performed on the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset. Then, target training data is constructed based on a preset time window and the target power-meteorological aligned dataset; the preset time window includes historical and forecast windows. Finally, the preset model is trained using the target training data to obtain a power prediction model. In this way, by supplementing the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset, the supplemented features help the preset model learn temporal relationships that cannot be reflected by power or meteorological data alone, improving the model's ability to perceive complex influencing factors within meteorological data. Furthermore, constructing the target training data based on the preset time window and the target power-meteorological aligned dataset improves the quality of the training data. In addition, training the preset model with the constructed target training data allows the preset model to learn how to predict future power generation based on historical and forecast meteorological data.

[0010] In some implementations, the actual power is collected by a first preset monitoring device at first preset time intervals; the measured meteorological data is collected by a second preset monitoring device at first preset time intervals.

[0011] In this way, the actual power is collected by the first preset monitoring device at first preset time intervals, and the measured meteorological data is collected by the second preset monitoring device at first preset time intervals. This ensures that the time granularity of the actual power and the measured meteorological data is consistent, improves the efficiency of subsequent data preprocessing, and strengthens the real-time correlation between the actual power and the measured meteorological data.

[0012] In some embodiments, the step of preprocessing the acquired actual power and measured meteorological data corresponding to the new energy station to determine the power-meteorological aligned dataset includes:

[0013] The actual power and the measured meteorological data are timestamped to obtain a first temporary power-meteorological aligned dataset;

[0014] Duplicate values ​​are removed from the first temporary power-meteorological aligned dataset to obtain the second temporary power-meteorological aligned dataset.

[0015] Outlier identification processing is performed on the second temporary power-meteorological aligned dataset to identify anomalous data;

[0016] Based on the missing data and the second temporary power-meteorological aligned dataset, the power-meteorological aligned dataset is determined, wherein the missing data includes the anomalous data and the uncollected actual power and / or the measured meteorological data.

[0017] Thus, timestamp alignment is performed on the actual power and measured meteorological data to obtain a first temporary power-meteorological aligned dataset. Next, duplicate values ​​are removed from the first temporary power-meteorological aligned dataset to obtain a second temporary power-meteorological aligned dataset. Then, outlier identification processing is performed on the second temporary power-meteorological aligned dataset to identify anomalous data. Finally, based on missing data and the second temporary power-meteorological aligned dataset, the final power-meteorological aligned dataset is determined. Missing data includes anomalous data and uncollected actual power and / or measured meteorological data. In this way, through data preprocessing, the final output "power-meteorological aligned dataset" possesses the characteristics of temporal consistency, unique information, and continuous completeness.

[0018] In some implementations, determining the power-weather aligned dataset based on the missing data and the second temporary power-weather aligned dataset includes:

[0019] The missing data that are consecutively missing and whose consecutive missing number is greater than or equal to a preset number are identified as consecutively missing sub-data.

[0020] From the second temporary power-meteorological aligned dataset, the corresponding time period subsequences corresponding to the continuous missing sub-data are removed to obtain the third temporary power-meteorological aligned dataset;

[0021] Based on a preset interpolation method, the missing data in the third temporary power-meteorological aligned dataset is filled to determine the power-meteorological aligned dataset.

[0022] Thus, consecutively missing data points with a number of consecutive missing data points greater than or equal to a preset number are identified as consecutively missing sub-data. Next, from the second temporary power-meteorological aligned dataset, the corresponding time-period subsequences of the consecutively missing sub-data are removed to obtain the third temporary power-meteorological aligned dataset. Finally, based on a preset imputation method, the missing data in the third temporary power-meteorological aligned dataset is filled to determine the power-meteorological aligned dataset. In this way, by removing consecutively missing sub-data and imputing the remaining small amount of missing data, invalid repair and noise introduction can be avoided, and the temporal continuity of the data can be maintained, thereby ensuring the quality of the power-meteorological aligned dataset.

[0023] In some implementations, the step of performing feature augmentation processing on the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset includes:

[0024] Based on the timestamp-derived features, feature supplementation processing is performed on the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset, wherein the timestamp-derived features are obtained based on standard timestamp splitting.

[0025] Thus, based on timestamp-derived features, feature supplementation processing is performed on the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset. The timestamp-derived features are obtained by splitting standard timestamps. In this way, through feature supplementation processing, abstract temporal information can be transformed into quantifiable and correlated structured features and combined with the power-meteorological aligned dataset. This allows the pre-defined model to understand how power changes over time, improving the accuracy of model predictions.

[0026] In some implementations, constructing target training data based on a preset time window and the target power-meteorological alignment dataset includes:

[0027] The target power-meteorological aligned dataset is stitched together to generate the first temporary training data;

[0028] Based on the preset time window, the first temporary training data is segmented to generate original training data fragments;

[0029] Based on a preset dimension, the original training data fragment is fused to generate an initial training data fragment. The preset dimension includes a power sequence, a meteorological feature sequence, and a time feature sequence.

[0030] Based on a preset mask matrix, the power sequence corresponding to the prediction window in the initial training data segment is masked to determine the target training data.

[0031] Thus, the target power-meteorological aligned dataset is stitched together to generate the first temporary training data. Next, based on a preset time window, the first temporary training data is segmented to generate original training data fragments. Then, based on preset dimensions, the original training data fragments are fused to generate initial training data fragments. The preset dimensions include power sequences, meteorological feature sequences, and time feature sequences. Finally, based on a preset mask matrix, the power sequences in the initial training data fragments corresponding to the prediction window are masked to determine the target training data. In this way, through the above stitching, segmentation, dimension fusion, and masking processes, the generated target training data structure is adapted and the information is complete, thereby providing high-quality input for the training of the preset model.

[0032] In some implementations, the preset model includes a block module and a temporal convolutional network block, wherein the temporal convolutional network block is composed of temporal mixing sub-blocks, feature mixing sub-blocks, variable mixing sub-blocks, and residual connections.

[0033] Thus, the pre-defined model includes a block module and a temporal convolutional network block. The temporal convolutional network block consists of temporal mixing sub-blocks, feature mixing sub-blocks, variable mixing sub-blocks, and residual connections. In this way, the block module can efficiently process long-term temporal data, balancing information preservation and computational efficiency. The temporal mixing sub-blocks enable the model to accurately grasp dynamic correlations along the time dimension, strengthening the learning of temporal patterns. The feature mixing and variable mixing sub-blocks enable the model to understand power changes from multiple dimensions. Residual connections ensure the model's training stability, thereby improving the model's fitting ability.

[0034] In some implementations, training the preset model based on the target training data includes:

[0035] Based on a preset loss algorithm, a preset model is trained using the target training data until the deviation between the predicted power obtained by the preset model and the actual power corresponding to the target training data is less than a preset threshold.

[0036] Thus, based on a preset loss algorithm, the preset model is trained using the target training data until the deviation between the predicted power obtained by the preset model and the actual power corresponding to the target training data is less than a preset threshold. This ensures that the model can stably and efficiently learn the changing patterns of new energy power, ultimately outputting a power prediction model with "high prediction accuracy and strong generalization ability".

[0037] This application provides a power prediction model, which is trained based on the power prediction model training method described in any one of claims 1-8.

[0038] This application provides an electronic device, which includes a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0039] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described above.

[0040] Additional aspects and advantages of embodiments of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of embodiments of this application. Attached Figure Description

[0041] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, wherein:

[0042] Figure 1 This is one of the flowcharts illustrating the training method of the power prediction model in certain embodiments of this application;

[0043] Figure 2 This is a second schematic flowchart of the training method for the power prediction model in some embodiments of this application;

[0044] Figure 3 This is the third flowchart illustrating the training method of the power prediction model in some embodiments of this application;

[0045] Figure 4 This is the fourth flowchart illustrating the training method of the power prediction model in some embodiments of this application;

[0046] Figure 5 This is the fifth flowchart illustrating the training method of the power prediction model in some embodiments of this application;

[0047] Figure 6 This is the sixth flowchart illustrating the training method of the power prediction model in some embodiments of this application. Detailed Implementation

[0048] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the embodiments of this application, and should not be construed as limiting the embodiments of this application.

[0049] Against the backdrop of an accelerated global energy structure transformation towards cleaner and lower-carbon energy sources, new energy sources (such as wind power and photovoltaic power) have become a core growth engine for the energy supply system. However, new energy power generation is characterized by strong intermittency, volatility, and randomness, and its output is highly coupled with meteorological conditions. For example, photovoltaic power depends on factors related to solar irradiance and temperature, while wind power is directly affected by airflow characteristics such as wind speed and direction at different altitudes. Therefore, accurately predicting the power generation of new energy sites under different meteorological data through power prediction models is a crucial support for ensuring stable grid operation, improving the absorption capacity of new energy sources, and optimizing electricity trading strategies.

[0050] However, existing power prediction models often rely on actual power and basic meteorological data during training, failing to make good use of future meteorological information obtained from weather forecasts. This results in a lack of ability to predict how future weather changes will affect power output, thus preventing existing power prediction models from effectively capturing the deep-seated patterns between new energy power generation and meteorological data, leading to consistently low prediction accuracy.

[0051] Based on the above issues, please refer to Figure 1 This application provides a method for training a power prediction model for a new energy power station, the method comprising:

[0052] 01: Perform data preprocessing on the actual power and measured meteorological data corresponding to the new energy stations to determine the power-meteorological aligned dataset;

[0053] 02: Perform feature supplementation processing on the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset;

[0054] 03: Construct target training data based on the preset time window and the target power-meteorological aligned dataset;

[0055] 04: Train the preset model based on the target training data to obtain the power prediction model.

[0056] This application also provides an electronic device, including a memory and a processor. The training method of the power prediction model of this application can be implemented by the electronic device of this application. Specifically, the memory stores a computer program, and the processor is used to perform data preprocessing on the actual power and measured meteorological data corresponding to the new energy station to determine the power-meteorological aligned dataset. It also performs feature supplementation processing on the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset. The processor is further used to construct target training data based on a preset time window and the target power-meteorological aligned dataset. Finally, it trains a preset model based on the target training data to obtain the power prediction model.

[0057] This application also provides a model training device. The training method of the power prediction model in this application can be implemented by the model training device. Specifically, the model training device includes a data preprocessing module, a feature supplementation module, a construction module, and a training module. The data preprocessing module is used to preprocess the actual power and measured meteorological data corresponding to the new energy stations to determine the power-meteorological aligned dataset. The feature supplementation module is used to supplement the features of the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset. The construction module is used to construct target training data according to a preset time window and the target power-meteorological aligned dataset. The training module is used to train the preset model according to the target training data to obtain the power prediction model.

[0058] Specifically, new energy sites refer to places or units that generate electricity using renewable energy sources (such as solar and wind power), including but not limited to centralized photovoltaic power plants, distributed photovoltaic areas, and wind farms.

[0059] Actual power refers to the actual electrical power output of a renewable energy power station at a specific historical moment, collected by the station's data acquisition and monitoring control system at a fixed frequency (e.g., every 15 minutes). Furthermore, actual power is the "label data" (i.e., the true value of the predicted target) used in model training, reflecting the station's actual power generation capacity under specific meteorological conditions.

[0060] Measured meteorological data refers to environmental meteorological parameters collected by meteorological monitoring equipment (such as anemometers, irradiance meters, and thermometers) deployed at new energy sites. Measured meteorological data includes conventional meteorological parameters such as temperature, humidity, air pressure, and precipitation, as well as specific parameters such as solar irradiance (direct irradiance and diffuse irradiance) at photovoltaic sites and wind speeds at various heights at wind farms.

[0061] Data preprocessing refers to the cleaning and standardization operations performed on the raw, collected actual power and measured meteorological data. Data preprocessing can eliminate noise, misalignment, and missing data in the raw data, ensuring that the data used for subsequent modeling is clean, complete, and time-consistent, thus avoiding the model learning errors caused by poor-quality data.

[0062] A power-meteorological aligned dataset refers to the direct output of data preprocessing. It is a structured dataset where each record contains a unique timestamp, the actual power value at that moment, and multi-dimensional measured meteorological parameters at that moment. For example, "2023-01-01 00:00" corresponds to a power value of "120kW", a temperature of "5℃", and a wind speed of "3m / s".

[0063] Feature enhancement processing refers to the process of enriching the information dimensions of the data by adding time-related derived features on the basis of the "power-meteorological aligned dataset" to help the model more accurately capture the temporal patterns and periodic characteristics of new energy power generation.

[0064] The target power-meteorological aligned dataset refers to an enhanced dataset formed by supplementing the "power-meteorological aligned dataset" with "timestamp-derived features." By introducing time features, the target power-meteorological aligned dataset transforms the "implicit temporal patterns" of new energy power generation into "explicit features that the model can recognize." For example, the diurnal periodicity of photovoltaic power (low in the morning, high at noon, low in the evening) is not only related to irradiance but also directly linked to the "hour"—by supplementing the "hour" feature, the model can learn the pattern of "power increasing synchronously when irradiance increases at 8 a.m." more quickly, improving its ability to capture temporal patterns.

[0065] The preset time window refers to the "time range parameter" used to segment training samples from the "target power-meteorological alignment dataset," including two key sub-windows: the historical window and the prediction window. The historical window refers to the length of historical data before the prediction time t, providing the model with "historical power change trends" and "historical meteorological evolution patterns," supporting future predictions. The prediction window refers to the length of future data to be predicted after the prediction time t.

[0066] The target training data refers to standardized training samples constructed according to a "preset time window" that can be directly input into the model. It should be noted that in related technologies, training data only uses "historical meteorological data" as input and cannot utilize future meteorological information. However, the training method provided in this application expands the input of training data to include "historical meteorological data + future meteorological forecast data," laying the foundation for incorporating numerical weather prediction during inference.

[0067] The pre-defined model refers to the deep learning network architecture used to learn the mapping relationship between "input features → output power".

[0068] After acquiring the raw actual power data and measured meteorological data from the renewable energy power generation sites, the data is preprocessed to obtain a power-meteorological aligned dataset. Next, feature enhancement processing is performed on the power-meteorological aligned dataset to transform the implicit temporal patterns of renewable energy power generation (such as seasonal / intra-day cycles) into explicit features that the model can recognize, thus strengthening the model's ability to learn from the combined effects of meteorological conditions and temporal patterns. Subsequently, target training data is constructed based on a preset time window and the target power-meteorological aligned dataset. Finally, the target training data is input into the preset model to train it, resulting in a power prediction model.

[0069] In summary, regarding the training method and electronic device for the power prediction model provided in this application, the electronic device performs data preprocessing on the actual power and measured meteorological data corresponding to the new energy station to determine a power-meteorological aligned dataset. Next, feature supplementation processing is performed on the power-meteorological aligned dataset to determine a target power-meteorological aligned dataset. Then, target training data is constructed based on a preset time window and the target power-meteorological aligned dataset, where the preset time window includes a historical window and a prediction window. Finally, the preset model is trained based on the target training data to obtain the power prediction model. Thus, by performing feature supplementation processing on the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset, the supplemented features can help the preset model learn temporal relationships that cannot be reflected by power or meteorological data alone, improving the model's ability to perceive complex influencing factors within the meteorological data. Furthermore, by constructing target training data based on the preset time window and the target power-meteorological aligned dataset, the quality of the training data is improved. In addition, training the preset model with the constructed target training data allows the preset model to learn how to predict future power generation based on historical and predicted meteorological data.

[0070] In some implementations, the actual power is collected by a first preset monitoring device at first preset time intervals; the measured meteorological data is collected by a second preset monitoring device at first preset time intervals.

[0071] Specifically, the first preset monitoring equipment refers to the hardware system used to collect actual power generation data from renewable energy power plants. For example, the data acquisition and monitoring control system is standard equipment at renewable energy power plants. This system can monitor the dynamic changes in power generation in real time, record and store data at a fixed frequency, and is the core equipment for monitoring the operation of renewable energy power plants.

[0072] The second set of pre-set monitoring equipment refers to a cluster of specialized instruments used to collect measured meteorological data from new energy power plants. Different equipment is configured according to the power generation type (photovoltaic / wind power). For example, the second set of pre-set monitoring equipment for photovoltaic power plants includes dedicated photovoltaic weather stations and general-purpose meteorological sensors.

[0073] The first preset duration refers to the unified data acquisition time interval between the first preset monitoring device and the second preset monitoring device, meaning that the data acquisition frequency of the two types of devices remains consistent. In the embodiment of this application, the first preset duration is "15 minutes". It should be noted that if the acquisition frequencies of the first preset monitoring device and the second preset monitoring device are inconsistent, it will cause the "power at a certain moment" to fail to match the "meteorological conditions of the same period", resulting in "time sequence misalignment". For example, "power at 8:15" may correspond to "meteorological data from 8:00 to 8:30", and the model will be unable to learn an accurate causal relationship.

[0074] In this way, the actual power is collected by the first preset monitoring device at first preset time intervals, and the measured meteorological data is collected by the second preset monitoring device at first preset time intervals. This ensures that the time granularity of the actual power and the measured meteorological data is consistent, improves the efficiency of subsequent data preprocessing, and strengthens the real-time correlation between the actual power and the measured meteorological data.

[0075] Please see Figure 2 In some implementations, step 01 (preprocessing the acquired actual power and measured meteorological data corresponding to the new energy station to determine the power-meteorological aligned dataset) includes:

[0076] 011: Timestamp alignment is performed on the actual power and measured meteorological data to obtain the first temporary power-meteorological aligned dataset;

[0077] 012: Duplicate values ​​are removed from the first temporary power-meteorological aligned dataset to obtain the second temporary power-meteorological aligned dataset;

[0078] 013: Perform outlier identification processing on the second temporary power-meteorological aligned dataset to determine anomalous data;

[0079] 014: Determine the power-weather aligned dataset based on the missing data and the second temporary power-weather aligned dataset.

[0080] In some implementations, the data preprocessing module is further used to perform timestamp alignment processing on the actual power and measured meteorological data to obtain a first temporary power-meteorological aligned dataset. It also performs duplicate value removal processing on the first temporary power-meteorological aligned dataset to obtain a second temporary power-meteorological aligned dataset. Finally, it performs outlier identification processing on the second temporary power-meteorological aligned dataset to determine abnormal data and, based on missing data and the second temporary power-meteorological aligned dataset, determines the power-meteorological aligned dataset.

[0081] In some implementations, the processor is further configured to perform timestamp alignment processing on the actual power and measured meteorological data to obtain a first temporary power-meteorological aligned dataset; and to perform duplicate value removal processing on the first temporary power-meteorological aligned dataset to obtain a second temporary power-meteorological aligned dataset; and to perform outlier identification processing on the second temporary power-meteorological aligned dataset to identify outlier data and, based on missing data and the second temporary power-meteorological aligned dataset, to determine the power-meteorological aligned dataset.

[0082] Specifically, timestamp alignment refers to the process of matching and binding the original collected "actual power data" and "measured meteorological data" one-to-one according to their collection timestamps (time markers accurate to minutes / seconds). This ensures that the "power-meteorological aligned dataset" is strictly synchronized in the time dimension, laying the foundation for "spatiotemporal matching" in subsequent processing. The steps of timestamp alignment are as follows: First, extract the timestamps of the actual power data and the measured meteorological data. Then, if there is a slight discrepancy between the timestamps of the actual power data and the measured meteorological data (e.g., power data marked 9:00, meteorological data marked 9:01), adjust them to the same time using a time calibration tool (e.g., uniformly convert to UTC time).

[0083] The first temporary power-meteorological aligned dataset refers to the direct output of "timestamp alignment processing," and is a preliminarily correlated structured dataset. Each data point in the first temporary power-meteorological aligned dataset includes a unique timestamp (e.g., "2024-07-01 09:00:00"), the actual power value corresponding to that timestamp, and the measured meteorological parameters corresponding to that timestamp (e.g., temperature 25℃, wind speed 4m / s, irradiance 800). ).

[0084] Duplicate value removal refers to the process of identifying and deleting multiple identical or highly similar records corresponding to the same timestamp from the "first temporary power-meteorological alignment dataset". In some implementations, the logic for determining duplicate values ​​may be that at the same timestamp, there are two or more records with completely identical power values ​​and meteorological parameters (e.g., both with a power of 850kW and a temperature of 25℃) or differences within the allowable error range (e.g., a power difference of less than or equal to 0.5kW and a temperature difference of less than or equal to 0.1℃).

[0085] The second temporary power-meteorological aligned dataset refers to the output of the "duplicate value removal process". It is an associated dataset after eliminating duplicates and redundancies. Its structure is consistent with the first temporary data (including timestamps, power, and meteorological parameters), but each timestamp corresponds to only one record.

[0086] Outlier identification and processing refers to the process of identifying anomalous records from the "second temporary power-meteorological aligned dataset" that do not conform to physical laws or statistical distributions. Outlier identification and processing is a "noise filtering" step in data preprocessing. Anomalous data is often generated due to sensor malfunctions (such as radiometer failure) or extreme interference (such as a sudden power surge caused by lightning strikes), and the "meteorological-power relationship" it reflects is spurious (e.g., power of 1000kW at 0 wind speed). If anomalous data is retained, the model will learn incorrect patterns (e.g., "high power generation is possible at 0 wind speed"), leading to decreased prediction accuracy. In some implementations, outlier identification and processing methods include statistical methods and physical methods. Statistical methods refer to methods based on the probability distribution of the data (e.g.,...). The standard method identifies outliers by considering the inherent characteristics of renewable energy generation. For example, at a photovoltaic site, the power output should be close to zero at night (irradiance = 0); a power output of 1600kW at a certain moment would be considered abnormal. The physical method involves combining the inherent laws of renewable energy generation to identify outliers.

[0087] Outlier data refers to the output of "outlier identification and processing", which refers to data that is determined to "not conform to physical laws or statistical distribution".

[0088] Missing data refers to "unrecorded or invalid" information in the second temporary power-meteorological aligned dataset, including outlier data (after outlier identification, outlier records are marked as missing values) and actual power and / or measured meteorological data that were not collected (original data gaps caused by equipment offline, communication interruption, sensor failure, etc., i.e. no power or meteorological records for a certain period of time).

[0089] Thus, timestamp alignment is performed on the actual power and measured meteorological data to obtain a first temporary power-meteorological aligned dataset. Next, duplicate values ​​are removed from the first temporary power-meteorological aligned dataset to obtain a second temporary power-meteorological aligned dataset. Then, outlier identification processing is performed on the second temporary power-meteorological aligned dataset to identify anomalous data. Finally, based on missing data and the second temporary power-meteorological aligned dataset, the final power-meteorological aligned dataset is determined. Missing data includes anomalous data and uncollected actual power and / or measured meteorological data. In this way, through data preprocessing, the final output "power-meteorological aligned dataset" possesses the characteristics of temporal consistency, unique information, and continuous completeness.

[0090] Please see Figure 3 In some implementations, step 014 (determining the power-weather aligned dataset based on the missing data and the second temporary power-weather aligned dataset) includes:

[0091] 0141: For missing data that are consecutively missing and whose consecutive missing number is greater than or equal to a preset number, determine the consecutively missing sub-data;

[0092] 0142: From the second temporary power-meteorological aligned dataset, remove the corresponding time period subsequences that correspond to the continuously missing sub-data to obtain the third temporary power-meteorological aligned dataset;

[0093] 0143: Based on the preset interpolation method, the missing data in the third temporary power-meteorological aligned dataset is filled to determine the power-meteorological aligned dataset.

[0094] In some implementations, the data preprocessing module is further configured to determine consecutively missing sub-data by identifying consecutively missing data with a number of consecutively missing data greater than or equal to a preset number; and to remove the corresponding time period subsequences from the second temporary power-meteorological aligned dataset to obtain a third temporary power-meteorological aligned dataset; and to fill in the missing data in the third temporary power-meteorological aligned dataset based on a preset interpolation method to determine the power-meteorological aligned dataset.

[0095] In some implementations, the processor is further configured to determine consecutively missing sub-data based on consecutively missing data with a number of consecutively missing data greater than or equal to a preset number; and to remove the corresponding time period subsequences corresponding to the consecutively missing sub-data from the second temporary power-meteorological aligned dataset to obtain a third temporary power-meteorological aligned dataset; and to fill in the missing data in the third temporary power-meteorological aligned dataset based on a preset interpolation method to determine the power-meteorological aligned dataset.

[0096] Specifically, consecutively missing data segments refer to fragments in the second temporary power-meteorological aligned dataset where "power or meteorological data for multiple consecutive timestamps are all missing values," and the number of consecutively missing data segments reaches or exceeds a preset number. In some implementations, the preset number is 16 timestamps. It should be noted that new energy power generation data has strong temporal correlation (e.g., power changes depend on preceding meteorological trends). If the consecutively missing time is too long (e.g., more than 4 hours), the correlation between the effective data before and after is weak, and the true trend cannot be accurately recovered through interpolation. Therefore, the consecutively missing data segments need to be removed.

[0097] Removing the corresponding time segment subsequence for consecutively missing sub-data refers to deleting the entire time segment containing the consecutively missing sub-data from the second temporary data. The specific operation may be as follows: First, locate the time range of the consecutively missing sub-data. Then, delete all records within that time range (regardless of whether other fields are missing) to ensure the continuity of the time sequence.

[0098] Preset imputation refers to a standardized method for filling sporadic missing values ​​(such as missing values ​​from 1 to 15 timestamps) in the third temporary data where "the number of consecutive missing values ​​is less than the preset number". In some implementations, preset imputation includes linear imputation and adjacent value imputation. Linear imputation calculates the imputation value based on the valid data before and after the missing value, following a linear trend over time. For example, if the power is 500kW at 8:00, 700kW at 8:30, and a missing value at 8:15, the imputation value is 500 + (700 - 500) × (15 / 30) = 600kW. Adjacent value imputation uses valid data from the time preceding the missing value (e.g., if 8:15 is missing, use the 500kW from 8:00), suitable for parameters with gradual changes (such as temperature).

[0099] Thus, consecutively missing data points with a number of consecutive missing data points greater than or equal to a preset number are identified as consecutively missing sub-data. Next, from the second temporary power-meteorological aligned dataset, the corresponding time-period subsequences of the consecutively missing sub-data are removed to obtain the third temporary power-meteorological aligned dataset. Finally, based on a preset imputation method, the missing data in the third temporary power-meteorological aligned dataset is filled to determine the power-meteorological aligned dataset. In this way, by removing consecutively missing sub-data and imputing the remaining small amount of missing data, invalid repair and noise introduction can be avoided, and the temporal continuity of the data can be maintained, thereby ensuring the quality of the power-meteorological aligned dataset.

[0100] Please see Figure 4 In some implementations, step 02 (performing feature supplementation on the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset) includes:

[0101] 021: Based on the timestamp-derived features, perform feature supplementation processing on the power-meteorological alignment dataset to determine the target power-meteorological alignment dataset.

[0102] In some implementations, the feature supplementation module is also used to perform feature supplementation processing on the power-meteorological aligned dataset based on features derived from timestamps, and to determine the target power-meteorological aligned dataset.

[0103] In some implementations, the processor is also configured to perform feature supplementation processing on the power-weather aligned dataset based on timestamp-derived features to determine the target power-weather aligned dataset.

[0104] Specifically, timestamp-derived features refer to structured temporal features related to the patterns of new energy power generation, extracted from the "standard timestamps" (such as "2024-08-01 09:15:00") in the power-meteorological alignment dataset. For example, the timestamp "2024-08-01 09:15:00" can be broken down into timestamp-derived features such as: hour=9, day of the week=4 (Thursday), month=8, quarter=3 (summer). Timestamp-derived features can transform "implicit temporal patterns" into "explicit features that the model can recognize." Specifically, new energy power generation has strong periodicity. For example, photovoltaic power varies with sunshine hours in a "single-peak curve" (low in the morning, high at noon, and low in the evening). There may be systematic differences in meteorological conditions between weekends and weekdays (such as local humidity changes due to increased industrial activity on weekdays), and high solar irradiance and peak photovoltaic power occur in summer. These strong periodicities cannot be directly reflected by "actual power" or "measured meteorological data" alone (for example, under the same irradiance, the photovoltaic power at 10:00 and 14:00 may differ due to different module temperatures). Timestamp-derived features, by quantifying time location, allow the model to capture "power differences under the same meteorological conditions at different times".

[0105] First, the timestamp-derived features corresponding to each timestamp (e.g., hour=9, month=8) are compared with the "actual power" and "measured meteorological parameters" under that timestamp (e.g., power 850kW, irradiance 900). The data (including temperature 26℃) is bound together to form a multi-dimensional record of "timestamp + power + meteorology + time feature". Then, the fused records are arranged in order of timestamp to form a new dataset - "target power-meteorology aligned dataset". Each record contains the original data (power, meteorology) and newly added time-derived features. The data dimension is expanded from "power + meteorology" to "power + meteorology + time feature".

[0106] Thus, based on timestamp-derived features, feature supplementation processing is performed on the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset. The timestamp-derived features are obtained by splitting standard timestamps. In this way, through feature supplementation processing, abstract temporal information can be transformed into quantifiable and correlated structured features and combined with the power-meteorological aligned dataset. This allows the pre-defined model to understand how power changes over time, improving the accuracy of model predictions.

[0107] Please see Figure 5 In some implementations, step 03 (constructing target training data based on a preset time window and a target power-meteorological alignment dataset) includes:

[0108] 031: The target power-meteorological aligned dataset is stitched together to generate the first temporary training data;

[0109] 032: Based on a preset time window, the first temporary training data is segmented to generate original training data fragments;

[0110] 033: Based on preset dimensions, perform dimensional fusion on the original training data fragments to generate initial training data fragments;

[0111] 034: Based on the preset mask matrix, the power sequence corresponding to the prediction window in the initial training data segment is masked to determine the target training data.

[0112] In some implementations, the construction module is further configured to stitch together the target power-meteorological aligned dataset to generate first temporary training data; and to segment the first temporary training data based on a preset time window to generate original training data segments; and to perform dimensionality fusion on the original training data segments based on a preset dimension to generate initial training data segments, and to mask the power sequences in the initial training data segments corresponding to the prediction window based on a preset mask matrix to determine the target training data.

[0113] In some implementations, the processor is further configured to stitch together the target power-meteorological aligned dataset to generate first temporary training data; and to segment the first temporary training data based on a preset time window to generate original training data segments; and to perform dimensionality fusion on the original training data segments based on a preset dimension to generate initial training data segments, and to mask the power sequences in the initial training data segments corresponding to the prediction window based on a preset mask matrix to determine the target training data.

[0114] Specifically, the stitching process refers to the horizontal stitching of "power, meteorological features, and time features" in the target power-meteorological aligned dataset according to feature dimensions. In particular, it merges the "actual power value," "multi-dimensional measured meteorological parameters" (such as temperature, irradiance, and wind speed), and "multi-dimensional time-derived features" (such as hour, month, and whether it is a peak period) from each record into a single vector containing all features at the same timestamp. For example, the original data for a certain timestamp might be: power 850kW, meteorological parameters (temperature 25℃, irradiance 900kJ / m²), etc. The time features (hour 9, month 8) are concatenated to form a vector: [850, 25, 900, 9, 8]. This concatenation process integrates scattered single-dimensional features into a multi-dimensional feature vector.

[0115] The first temporary training data refers to the output of the concatenation process, which is a multi-dimensional feature sequence arranged in timestamp order. Each record is a vector containing "power + meteorological + time features", forming a time-series matrix (rows represent timestamps, and columns represent feature dimensions). For example, if the target power-meteorological aligned dataset contains 1000 timestamps, each timestamp corresponding to 5 features (1 power + 2 meteorological + 2 time features), then the first temporary training data is a 1000×5 matrix.

[0116] The preset time window refers to the time range parameter used to extract samples from the first temporary training data, including the historical window and the prediction window. The historical window *w* refers to the number of past timestamps to be included in the sample input (e.g., *w=64*, corresponding to 64 past 15-minute intervals, or 16 hours). The prediction window *h* refers to the number of future timestamps to be predicted from the sample output (e.g., *h=16*, corresponding to 16 future 15-minute intervals, or 4 hours).

[0117] Segmentation refers to the operation of sliding and extracting continuous time-series segments from the first temporary training data according to a preset time window, which can transform long time-series data into multiple independent sample segments. By sliding the preset window, each sample contains a complete range of "historical input + future prediction", ensuring that the model can learn the temporal patterns at different time points.

[0118] The original training data segments are the output of the segmentation process, which are time series subsequences corresponding to each preset time window. Each segment contains complete features (power, meteorological, and time features) for w+h timestamps. For example, the historical window part consists of 64 timestamps from tw to t-1 (including power, meteorological, and time features), and the prediction window part consists of 16 timestamps from t to t+h-1 (including power, meteorological, and time features).

[0119] Preset dimensions refer to the three main categories into which features in the original training data segments are divided according to their physical meaning. These categories are used to define the feature composition of the model input, including power sequences, meteorological feature sequences, and temporal feature sequences. The power sequence refers to the sequence composed of actual power values ​​at all timestamps (e.g., 64 power values ​​from the historical window + 16 power values ​​from the prediction window). The meteorological feature sequence refers to a multidimensional sequence composed of measured meteorological parameters (e.g., temperature, irradiance, wind speed) at all timestamps (each timestamp contains multiple meteorological features). The temporal feature sequence refers to a multidimensional sequence composed of time-derived features (e.g., hour, month, peak / peak status) at all timestamps (each timestamp contains multiple temporal features).

[0120] Dimension fusion is an operation that splits the original training data segments according to preset dimensions and then reassembles them along the feature dimensions. In some implementations, the steps of dimension fusion may be as follows: split the original segments according to "power sequence, meteorological feature sequence, and time feature sequence". Then, concatenate the three along the feature dimensions (e.g., the power sequence is 1-dimensional, the meteorological sequence is 5-dimensional, and the time series is 3-dimensional, resulting in 9 dimensions after fusion) to form a multi-dimensional time series matrix of "(w+h)×(1+5+3)".

[0121] The initial training data fragments are the output of dimension fusion, which are multi-dimensional time series matrices integrated according to preset dimensions. Its structure is as follows: each fragment contains w+h timestamps, each timestamp corresponds to the fusion feature of "power + weather + time" (such as 9-dimensional features), and the range of "historical window (input)" and "prediction window (output)" are clearly defined.

[0122] The preset mask matrix is ​​a 0-1 matrix with the same dimensions as the initial training data fragment, used to mask the power sequence within the prediction window. The construction rules for the preset mask matrix are as follows: the corresponding positions of the power sequence in the historical window (tw to t-1) are 1 (reserved); the corresponding positions of the power sequence in the prediction window (t to t+h-1) are 0 (masked); all positions of the meteorological feature sequence and the time feature sequence are 1 (reserved, because future meteorological data is needed for prediction).

[0123] Masking is an operation that multiplies the initial training data segment element by element with a preset mask matrix, so that the power sequence within the prediction window is set to 0 (masked), while the other features (historical power, all-time weather and time features) remain unchanged.

[0124] The target training data is the final output of the masking process and is a standardized sample set that can be directly input into the model.

[0125] Thus, the target power-meteorological aligned dataset is stitched together to generate the first temporary training data. Next, based on a preset time window, the first temporary training data is segmented to generate original training data fragments. Then, based on preset dimensions, the original training data fragments are fused to generate initial training data fragments. The preset dimensions include power sequences, meteorological feature sequences, and time feature sequences. Finally, based on a preset mask matrix, the power sequences in the initial training data fragments corresponding to the prediction window are masked to determine the target training data. In this way, through the above stitching, segmentation, dimension fusion, and masking processes, the generated target training data structure is adapted and the information is complete, thereby providing high-quality input for the training of the preset model.

[0126] In some implementations, the preset model includes a block module and a temporal convolutional network block, which consists of a temporal mixing sub-block, a feature mixing sub-block, a variable mixing sub-block, and residual connections.

[0127] Specifically, the block-based module refers to the "front-end preprocessing unit" of the pre-defined model. Through convolutional operations with large kernels and large strides, it divides the input long-term target training data (such as time-series data with "historical windows + prediction windows") into several "time patches" along the time dimension, and aggregates the feature information within each patch. For example, for a time-series sequence of length 80 (64 historical + 16 predictions), using a convolutional layer with a kernel size of 8 and a stride of 4, the sequence can be divided into (80-8) / 4+1=19 time patches. Each patch corresponds to the information of 8 consecutive timestamps in the original sequence, and after convolution, 19 "patch feature vectors" are output (each vector integrates the information of the 8 timestamps within the patch). Thus, since new energy power prediction requires processing historical data of several hours or even days (e.g., 16 hours = 64 points), directly inputting long-term raw data into the model would significantly increase the number of parameters and computational complexity. Through patch aggregation, the pre-defined model can reduce the data dimensionality while retaining key time-series information.

[0128] Temporal convolutional network blocks extract deep temporal correlations and feature interactions from the input data by stacking multiple network blocks with the same structure (usually 3-5 layers). Each block consists of four parts: "temporal mixture sub-block, feature mixture sub-block, variable mixture sub-block, and residual connection", forming a progressive learning process of "temporal → single feature → multiple features".

[0129] The temporal hybrid sub-block combines depthwise separable convolution and dilated convolution to capture the impact of historical information over a long time span on current predictions. This means it learns the "impact of time on features," adapting to the strong temporal nature of new energy power generation. Depthwise separable convolution is used to convolve each feature channel of the input independently (e.g., performing temporal convolution on features like "temperature" and "irradiance"), avoiding cross-feature interference and focusing on learning the temporal changes of a single feature (e.g., the rise and fall trend of temperature over time). Dilated convolution expands the receptive field by introducing "holes" (gaps) into the convolution kernel (e.g., a 3×1 convolution kernel with a dilation rate of 2 is equivalent to a 5×1 ordinary convolution kernel), capturing correlations at more distant time points without increasing the number of parameters.

[0130] Feature mixing sub-blocks are implemented through grouped point convolutions, focusing on nonlinear relationships within the same feature (such as the nonlinear conversion relationship between irradiance and power). Grouped point convolutions divide the channels of a single feature into groups (e.g., dividing the 64 feature channels of "irradiance" into 8 groups), performing 1×1 convolutions (point convolutions) within each group. This enhances the nonlinear expressive power of the feature (through activation functions such as ReLU) while reducing computational cost (the number of parameters is reduced to 1 / 8 compared to a standard 1×1 convolution). In renewable energy generation, many features exhibit nonlinear relationships (e.g., photovoltaic power and irradiance saturate in high-value regions due to increased module temperature, no longer increasing linearly). Feature mixing sub-blocks, through grouped convolutions and nonlinear activations, allow the model to learn these "nonlinear laws within a single feature," avoiding prediction biases caused by simple linear assumptions.

[0131] The variable mixing sub-block also uses grouped point convolution, but focuses on the interaction between different variables (such as power, temperature, irradiance, and time characteristics) (e.g., the combined suppression effect of "high irradiance + high temperature" on photovoltaic power). The grouped point convolution in the variable mixing sub-block mixes the feature channels of all variables in groups (e.g., merging the channels of features such as "temperature," "irradiance," and "hour" before grouping them), and performs a 1×1 convolution within each group to capture cross-variable correlations (e.g., "hour=12" (noon) and "irradiance=1000"). The combination of these factors amplifies the power output. Since renewable energy power is influenced by multiple factors including weather and time (e.g., a 5m / s wind speed in winter corresponds to higher wind power output than a 5m / s wind speed in summer due to different air densities), a single feature cannot explain these patterns. Variable mixing, through cross-variable convolution, allows the model to learn the mapping from "feature combination → power output," improving its adaptability to complex scenarios.

[0132] Residual connections refer to the element-wise addition of the target training data input to a temporal convolutional network block with the output features processed by three sub-blocks: temporal mixing, feature mixing, and variable mixing. This forms a closed loop of "input → processing → residual fusion." Specifically, when a model stacks multiple layers of temporal convolutional network blocks, the gradient may decay to zero during backpropagation due to multiple multiplications, making deep networks difficult to train. Residual connections directly retain the input feature x, allowing the gradient to propagate via a "shortcut," ensuring effective learning for deep networks. Simultaneously, the "x + F(x)" form allows the model to focus on learning the "difference between input and output" (i.e., F(x)), reducing the learning difficulty.

[0133] This modular design allows the model to accurately capture the temporal patterns of new energy power generation and integrate the complex correlations between multiple source features, ultimately achieving high-precision power prediction. Furthermore, because it learns the complex relationships between different features, the resulting power prediction model has strong transferability.

[0134] Thus, the pre-defined model includes a block module and a temporal convolutional network block. The temporal convolutional network block consists of temporal mixing sub-blocks, feature mixing sub-blocks, variable mixing sub-blocks, and residual connections. In this way, the block module can efficiently process long-term temporal data, balancing information preservation and computational efficiency. The temporal mixing sub-blocks enable the model to accurately grasp dynamic correlations along the time dimension, strengthening the learning of temporal patterns. The feature mixing and variable mixing sub-blocks enable the model to understand power changes from multiple dimensions. Residual connections ensure the model's training stability, thereby improving the model's fitting ability.

[0135] Please see Figure 6 In some implementations, step 04 (training the preset model based on the target training data) includes:

[0136] 041: Based on the preset loss algorithm, the preset model is trained according to the target training data until the deviation between the predicted power obtained by the preset model and the actual power corresponding to the target training data is less than the preset threshold.

[0137] In some implementations, the training module is used to train a preset model based on a preset loss algorithm and target training data until the deviation between the predicted power obtained by the preset model and the actual power corresponding to the target training data is less than a preset threshold.

[0138] In some implementations, the processor is further configured to train a preset model based on a preset loss algorithm and target training data until the deviation between the predicted power obtained by the preset model and the actual power corresponding to the target training data is less than a preset threshold.

[0139] Specifically, the preset loss algorithm refers to a mathematical formula for calculating the degree of difference between the predicted power and the actual power of the model. In some implementations, the preset loss algorithm may be the mean squared error loss.

[0140] The deviation value refers to the calculation result of the preset loss algorithm, and its physical meaning is "the average level of model prediction error".

[0141] The preset threshold refers to the "pass line" set according to the prediction accuracy requirements of the actual application scenario, and can be adjusted according to the actual situation.

[0142] Thus, based on a preset loss algorithm, the preset model is trained using the target training data until the deviation between the predicted power obtained by the preset model and the actual power corresponding to the target training data is less than a preset threshold. This ensures that the model can stably and efficiently learn the changing patterns of new energy power, ultimately outputting a power prediction model with "high prediction accuracy and strong generalization ability".

[0143] This application also provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program implements the steps of the training method for the power prediction model as described above.

[0144] It is understood that a computer program includes computer program code. Computer program code can be in the form of source code, object code, executable files, or some intermediate form. Computer-readable storage media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), and software distribution media, etc.

[0145] In this specification, the terms "specifically," "furthermore," "particularly," "understandably," etc., refer to specific features, structures, materials, or characteristics described in connection with embodiments or examples that are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0146] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of executable request code comprising one or more steps for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0147] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A training method for a power prediction model of a new energy power station, characterized in that, The method includes: Data preprocessing is performed on the actual power and measured meteorological data corresponding to the new energy stations to determine the power-meteorological aligned dataset; Based on timestamp-derived features, feature supplementation processing is performed on the power-meteorological aligned dataset to determine the target power-meteorological aligned dataset, wherein the timestamp-derived features are obtained based on standard timestamp splitting; Based on a preset time window and the target power-meteorological aligned dataset, target training data is constructed, wherein the preset time window includes a historical window and a prediction window; The preset model is trained based on the target training data to obtain the power prediction model; The step of constructing target training data based on a preset time window and the target power-meteorological aligned dataset includes: The target power-meteorological aligned dataset is stitched together to generate the first temporary training data; Based on the preset time window, the first temporary training data is segmented to generate original training data fragments; Based on a preset dimension, the original training data fragment is fused to generate an initial training data fragment. The preset dimension includes a power sequence, a meteorological feature sequence, and a time feature sequence. Based on a preset mask matrix, the power sequence corresponding to the prediction window in the initial training data segment is masked to determine the target training data.

2. The method according to claim 1, characterized in that, The actual power is collected by the first preset monitoring device at a first preset time interval; the measured meteorological data is collected by the second preset monitoring device at the first preset time interval.

3. The method according to claim 2, characterized in that, The process of preprocessing the acquired actual power and measured meteorological data corresponding to the new energy station to determine the power-meteorological aligned dataset includes: The actual power and the measured meteorological data are timestamped to obtain a first temporary power-meteorological aligned dataset; Duplicate values ​​are removed from the first temporary power-meteorological aligned dataset to obtain the second temporary power-meteorological aligned dataset. Anomaly identification processing was performed on the second temporary power-meteorological aligned dataset to identify anomalous data; Based on the missing data and the second temporary power-meteorological aligned dataset, the power-meteorological aligned dataset is determined, wherein the missing data includes the anomalous data and the uncollected actual power and / or the measured meteorological data.

4. The method according to claim 3, characterized in that, The step of determining the power-meteorological aligned dataset based on the missing data and the second temporary power-meteorological aligned dataset includes: The missing data that are consecutively missing and whose consecutive missing number is greater than or equal to a preset number are identified as consecutively missing sub-data. From the second temporary power-meteorological aligned dataset, the corresponding time period subsequences corresponding to the continuous missing sub-data are removed to obtain the third temporary power-meteorological aligned dataset; Based on a preset interpolation method, the missing data in the third temporary power-meteorological aligned dataset is filled to determine the power-meteorological aligned dataset.

5. The method according to claim 1, characterized in that, The preset model includes a block module and a temporal convolutional network block, which is composed of temporal mixing sub-blocks, feature mixing sub-blocks, variable mixing sub-blocks and residual connections.

6. The method according to claim 1, characterized in that, The step of training the preset model based on the target training data includes: Based on a preset loss algorithm, a preset model is trained using the target training data until the deviation between the predicted power obtained by the preset model and the actual power corresponding to the target training data is less than a preset threshold.

7. A power prediction model, characterized in that, The power prediction model is trained based on the power prediction model training method described in any one of claims 1-6.

8. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-6.

Images