System and method for predicting generated power and generated energy of wind power plant

Abstract

The invention relates to the technical field of electric power, in particular to a wind power plant generation power and generating capacity prediction system and method, and the system comprises a multi-source data collection module which is used for collecting wind turbine generator operation data, in-field wind measurement data, short-term / medium-term / long-term meteorological data, extreme weather early warning data, power grid load data and wind power plant environment data; an intelligent data preprocessing module; a multi-scale interaction modeling module; a dynamic feature weighting module; a real-time correction module; and a prediction result fusion and output module. According to the invention, comprehensive acquisition is realized through the multi-source data acquisition module, the cross-scale prediction deviation is reduced to be less than 6% in combination with the'anomaly detection-deletion repair-standardization 'whole-process processing of the intelligent preprocessing module, short-term, medium-term and long-term prediction results form a logic closed loop, and the prediction accuracy is improved. And the multi-level requirements of power grid dispatching on short-term accurate response, medium-term planning and long-term planning reference are met.

Description

Technical Field

[0001] This invention relates to the field of power technology, specifically to a wind farm power generation and electricity generation prediction system and a wind farm power generation and electricity generation prediction method. Background Technology

[0002] With the acceleration of the global energy transition, wind power, as a core component of clean and renewable energy, has seen its installed capacity continue to grow. However, the intermittency, volatility, and randomness of wind energy, coupled with the fact that the wind farm power generation process is affected by multiple factors such as "equipment status, meteorological conditions, grid load, and extreme weather", have made power generation and power prediction a key bottleneck restricting the large-scale grid connection of wind power, grid load balance, and efficient operation of wind farms. Summary of the Invention

[0003] This invention addresses the problems in the prior art by providing a system and method for predicting wind farm power generation and output. The specific technical solution is as follows: On the one hand, this application provides a wind farm power generation and electricity generation prediction system, including: The multi-source data acquisition module is used to collect wind turbine operation data, on-site wind measurement data, short-term / medium-term / long-term meteorological data, extreme weather warning data, power grid load data, and wind farm environmental data. The intelligent data preprocessing module is used to perform anomaly detection and removal, missing value repair, and standardization processing on the collected data; The multi-scale interactive modeling module includes a short-term sub-model, a medium-term sub-model, a long-term sub-model, and a multi-scale interactive unit. The short-term sub-model adopts Bi-LSTM, the medium-term sub-model adopts XGBoost, the long-term sub-model adopts an improved Transformer, and the multi-scale interactive unit is used to realize parameter transfer and result correction between sub-models. The dynamic feature weighting module is used to adjust feature weights in real time based on scene recognition results through an attention mechanism; The real-time correction module is used to calculate the prediction bias using a rolling window mechanism, and to fine-tune the model parameters and correct the prediction results based on the bias. The prediction result fusion and output module is used to fuse the results of multiple sub-models based on confidence weights and output the predicted values ​​and uncertainty intervals.

[0004] As a further technical solution of the present invention, the real-time sensor data sampling frequency of the multi-source data acquisition module is 10Hz, and the meteorological data update frequency is: 15 minutes / time for short-term, 1 hour / time for medium-term, and 1 day / time for long-term. The multi-source data acquisition module includes an extreme weather warning API interface.

[0005] As a further technical solution of the present invention, the intelligent data preprocessing module uses the isolated forest algorithm to detect abnormal data, uses the improved WGAN-GP to complete missing data, uses Z-score normalization to process numerical features, and uses one-hot encoding to process categorical features.

[0006] As a further technical solution of the present invention, the working mode of the multi-scale interaction unit includes: adjusting the feature weights of the intermediate sub-model with the trend coefficient output by the long-term sub-model, correcting the input features of the short-term sub-model with the fluctuation coefficient output by the intermediate sub-model, and adjusting the trend prediction of the long-term sub-model with the deviation feedback output by the short-term sub-model.

[0007] As a further technical solution of the present invention, the dynamic feature weighting module includes a scene recognition unit, an attention weight calculation unit and a weight update unit. The scene recognition unit can recognize scenes such as "normal weather", "typhoon warning" and "cold wave". The attention weight calculation unit uses the softmax function to calculate the feature weights. The weight update unit updates the weights every 30 minutes.

[0008] As a further technical solution of the present invention, the rolling window duration of the real-time correction module is 1 hour, the output layer parameters of the sub-model are fine-tuned using Mini-batchSGD, and the deviation value is superimposed on the prediction result of the next window.

[0009] As a further technical solution of the present invention, the confidence level of the prediction result fusion and output module is calculated based on the prediction error of the sub-model in the last 10 predictions, and the uncertainty range is "predicted value ± 2 × standard deviation of the error in the last 10 predictions".

[0010] On the other hand, this application also provides a method for predicting the power generation and electricity output of a wind farm, including the following steps: S1: Collect multi-dimensional data through a multi-source data acquisition module; S2: The intelligent data preprocessing module performs anomaly removal, missing data repair, and standardization. S3: Through the multi-scale interactive modeling module, the preprocessed data is input into the short-term, medium-term and long-term sub-models, and the prediction results are corrected through the multi-scale interactive unit to obtain P1, P2 and P3; S4: Through the dynamic feature weighting module, adjust the feature weights based on scene recognition and update P1, P2, and P3; S5: Through the real-time correction module, the deviation is calculated using a rolling window, the model parameters are fine-tuned and the prediction results are corrected to obtain P1', P2', and P3'. S6: Through the prediction result fusion and output module, P1', P2', and P3' are fused based on confidence weights to output the final predicted value and uncertainty range.

[0011] As a further technical solution of the present invention, the short-term sub-model is Bi-LSTM, which takes data from the next 2 hours as input; the medium-term sub-model is XGBoost, which takes data from the next 7 days as input; and the long-term sub-model is an improved Transformer, which takes data from the next 1 year as input.

[0012] As a further technical solution of the present invention, in step S6, the formula for calculating the final predicted value P is: C1, C2, and C3 represent the confidence levels of the short-term, medium-term, and long-term sub-models, respectively, and the confidence level is the average of the prediction errors from 1 to the last 10 predictions.

[0013] The beneficial effects of this invention are as follows: The system comprehensively collects 10Hz high-frequency wind turbine operation data, 1Hz field wind measurement data, multi-frequency meteorological data, and extreme weather warning data through a multi-source data acquisition module. Combined with the intelligent preprocessing module's "anomaly detection-missing data repair-standardization" full-process processing: it uses the isolated forest algorithm to remove abnormal data, uses the improved WGAN-GP to complete missing data, and uses Z-score normalization and one-hot encoding to unify the data format, reducing the cross-scale prediction deviation to below 6%. The short-term, medium-term, and long-term prediction results form a logical closed loop, meeting the multi-level needs of power grid dispatch for "accurate short-term response, medium-term planning, and long-term planning reference".

[0014] This scheme constructs a dynamic feedback mechanism for sub-models through multi-scale interactive units: the quarterly wind speed trend coefficient output by the long-term sub-model can adjust the daily average wind speed characteristic weight of the medium-term sub-model; the weekly load fluctuation coefficient output by the medium-term sub-model can correct the grid load input of the short-term sub-model; and the real-time deviation of the short-term sub-model can be fed back to optimize the equipment aging coefficient of the long-term sub-model, thus meeting the multi-level requirements of grid dispatch for "precise short-term response, medium-term planning, and long-term planning reference".

[0015] This solution achieves scene adaptive adjustment through a dynamic feature weighting module: the scene recognition unit can respond to extreme warnings within 1 minute, and the attention weight calculation unit increases the weights of key features such as wind direction stability and typhoon path distance to 0.38 and 0.19 respectively through the softmax function, and updates the weights and scene switching in real time every 30 minutes.

[0016] The system achieves dynamic compensation for deviations through a real-time correction module: when the deviation of two consecutive windows is ≤1%, the prediction value for the next period is adjusted in reverse by a correction formula so that the cumulative prediction deviation within 24 hours is ≤0.5MW; at the same time, the prediction result fusion module calculates the confidence level based on the error of the sub-model in the last 10 times, and further improves the accuracy through weighted fusion, and outputs the uncertainty range of "predicted value ±2 × error standard deviation". Attached Figure Description

[0017] Figure 1 This is a structural diagram of a wind farm power generation and electricity generation prediction system. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments.

[0019] This embodiment uses a large wind farm in the southeast coastal area as the application scenario, with a total installed capacity of 150MW, including 30 offshore wind turbines with a single unit capacity of 5MW. There are 6 70-meter-high wind measurement towers in the wind farm, located in the east, south, west, north, central and edge areas of the wind farm. The farm is affected by 2-3 typhoons per year. The grid connection point is the local 220kV substation, which mainly undertakes the task of supplementing the power supply for regional industrial and residential electricity loads.

[0020] The prediction period and time resolution of this embodiment are as follows: Short-term forecast: The next 2 hours, with a time resolution of 15 minutes; that is, one power prediction value is output every 15 minutes. Medium-term forecast: 7 days from now, with a time resolution of 1 hour; that is, one power prediction value is output every hour, accumulating 24 values ​​per day; Long-term forecast: 1 year from now, with a time resolution of 1 day; that is, outputting 1 daily power generation forecast value every day.

[0021] The data sources are as follows: Wind turbine operating data: collected through the unit's SCADA system, including speed, torque, gearbox temperature, generator stator temperature, blade pitch angle and fault warning signals, with a sampling frequency of 10Hz; On-site wind measurement data: collected by ultrasonic anemometers from 6 wind measurement towers, including wind speed, wind direction, turbulence intensity, and atmospheric pressure, with a sampling frequency of 1Hz; Meteorological data: Short-term meteorological data for the next 2 hours comes from the millimeter-wave radar built by the wind farm; medium-term meteorological data for the next 7 days comes from the Meteorological Bureau's Public Meteorological Service API; and long-term meteorological data for the next year comes from the annual climate trend forecast report released by the Meteorological Bureau's Climate Center. Extreme weather warning data: obtained through the Emergency Management Department's "Typhoon Warning API Interface", including typhoon number, central pressure, maximum wind speed, movement path, and warning level; Grid load data: Real-time load monitoring system from the power grid dispatch center, including regional total load demand and wind power grid connection quota, updated every 5 minutes; Wind farm environmental data includes the wind farm's latitude and longitude, seabed topography slope, annual average seawater temperature, and historical equipment operation and maintenance records.

[0022] This system adopts an "edge computing and cloud deployment" architecture. The edge is responsible for real-time data acquisition and preprocessing, while the cloud is responsible for modeling, prediction, and result output. The specific deployment and parameter settings are as follows: Deployment of multi-source data acquisition modules Hardware equipment: An edge computing gateway is adopted. Each gateway connects to the SCADA data of 5 wind turbines and the data of 1 wind measurement tower. The gateway supports 4G / 5G dual-mode communication with an uplink bandwidth of ≥100Mbps to ensure real-time transmission of 10Hz sampled data. Data transmission protocols: Wind turbine operation data is retrieved using the MQTT protocol, while meteorological and power grid data are retrieved using the HTTP / HTTPS protocol; Update frequency control: Short-term meteorological data sends a data request to the millimeter-wave radar every 15 minutes; medium-term meteorological data calls the meteorological bureau API once every hour; long-term meteorological data automatically downloads the climate trend report at 00:00 every day; extreme weather warning data adopts a dual mechanism of "push and pull". Data storage: Local SSD storage is used at the edge to store nearly 72 hours of raw data; Alibaba Cloud OSS object storage is used in the cloud to store 10 years of pre-processed data and supports fast query by timestamp.

[0023] Intelligent data preprocessing module parameters Anomaly detection uses the Isolation Forest algorithm: the number of trees is set to 100, and the sample subset size is set to 256. Outlier detection threshold: When the outlier score of a sample is ≥0.8, it is considered outlier data; for example, if the wind speed of a certain wind measuring tower suddenly increases from 5m / s to 28m / s within 10 seconds and there is no extreme weather warning, it is considered sensor fault data.

[0024] Missing value repair: Generator network structure: Input layer → Hidden layer 1 → Hidden layer 2 → Output layer; Discriminator network structure: Input layer → Hidden layer 1 → Hidden layer 2 → Output layer; Training parameters: batch size = 32, learning rate = 0.0001, training epochs = 100, gradient penalty coefficient = 10.

[0025] Data standardization: Numerical features such as wind speed and temperature are normalized using Z-score, the formula is as follows: , in the formula This represents the average value of this characteristic over the past 30 days. This represents the standard deviation of this characteristic over the past 30 days; for example, wind speed. 8.5m / s The value is 2.3 m / s, and the result after normalization for a wind speed of 10 m / s is approximately 0.65; Category features such as equipment operating status are encoded using one-hot encoding. Equipment operating status includes "normal", "maintenance" and "fault". After encoding, the dimension is 3, with "normal" corresponding to [1,0,0], "maintenance" corresponding to [0,1,0] and "fault" corresponding to [0,0,1].

[0026] Multi-scale interactive modeling module parameters Short-term sub-model: Input feature dimensions: 15, including 15 features such as wind speed, wind direction, rotational speed, torque, gearbox temperature, power grid load, and equipment status; Network structure: Input layer → Embedded layer → Bi-LSTM layer → Fully connected layer → Output layer; Training parameters: The optimizer was set to Adam, the learning rate was set to 0.001, the loss function was set to mean squared error, and the training data consisted of 15-minute resolution data from the past year, with a sample size of approximately 35,040. Medium-term sub-model: Input feature dimensions: 12, including 12 features such as daily average wind speed, daily maximum wind speed, seasonal factors, weekday / weekend identifier, and daily peak load of the power grid. Model parameters: number of trees set to 100, maximum tree depth set to 8, learning rate set to 0.1, minimum number of sample splits set to 5, loss function set to squared regression error; Training data: 1-hour resolution data from the past 3 years, with a sample size of approximately 26,280; Long-term sub-model: Input feature dimensions: 8, including 8 features such as monthly average wind speed, seasonal trend factor, equipment aging coefficient, annual maintenance frequency, and annual power grid load forecast; Network structure: Input layer → Position coding layer → Encoder → Fully connected layer → Output layer; Training parameters: The optimizer was set to AdamW, the learning rate was set to 0.0001, the weight decay was set to 0.01, and the training data consisted of daily resolution data from the past 5 years, with a sample size of approximately 1825.

[0027] Multi-scale interactive unit: Trend coefficient calculation: The long-term sub-model outputs the "quarterly wind speed trend coefficient". The calculation method is to divide the current quarter's average wind speed forecast by the annual average wind speed; Fluctuation coefficient calculation: The medium-term sub-model outputs the "weekly load fluctuation coefficient". The calculation method is to divide the current weekly peak load by the monthly average peak load. Bias Feedback: Short-term sub-model outputs "real-time bias" The calculation method is to subtract the absolute value of the actual power from the absolute value of the short-term predicted power. When the actual power is reached, the aging coefficient of the long-term sub-model equipment is updated.

[0028] Dynamic feature weighting module parameters Scene recognition unit: Scene classification thresholds: When the typhoon warning level is ≥ blue, it is judged as "typhoon scene", corresponding to a maximum wind speed at the center ≥ 10.8 m / s; when the temperature is ≤ 0℃ and the duration is ≥ 6 hours, it is judged as "cold wave scene"; when there is no extreme warning and the wind speed is in the range of 5-12 m / s, it is judged as "normal weather scene". Scene trigger delay: The delay from receiving the warning signal to scene switching is ≤1 minute.

[0029] Attention weight calculation unit: Feature importance score Obtained through feature importance evaluation using XGBoost; for example, wind direction stability. typhoon path distance Equipment wind resistance threshold Other features ; Weight calculation formula: For example, wind direction stability weights .

[0030] Weight update unit: It adopts a dual mechanism of "timed + event triggered", updating the weight once every 30 minutes. At the same time, when the scene changes from "normal" to "typhoon", the weight update is triggered immediately.

[0031] Real-time calibration module parameters Scrolling window settings: Set the window duration to 1 hour and the window scrolling step size to 15 minutes, that is, a new window is generated every 15 minutes. Typical windows include 14:00-15:00, 14:15-15:15, etc. Mini-batchSGD parameters: batch size set to 32, learning rate set to 0.0005, momentum set to 0.9, fine-tuning parameter is the sub-model output layer weight matrix, a typical example is Bi-LSTM output layer weight dimension of 32×1.

[0032] Deviation correction logic: Let the current window prediction value be... The measured value is ,deviation The formula for correcting the next window prediction is: ,in For correction factor, Set it to 0.8 to avoid over-correction.

[0033] Prediction result fusion and output module parameters Confidence calculation: Absolute error of the statistical sub-model in the last 10 predictions The calculation method is to subtract the absolute value of the actual power from the absolute value of the predicted power each time; Calculate the mean error The formula is ; The confidence level C is calculated using the following formula: ,in The average of the last 10 measured values ​​is used to ensure that C is between 0 and 1. Uncertainty interval calculation: Calculate the standard deviation of the errors over the last 10 times. The formula is ; The formula for calculating the uncertainty interval is: , of which 2 Corresponding to a 95% confidence level; Output format: The cloud-based web visualization platform outputs the prediction results, including short-term power curves, medium-term power heat maps, and long-term power generation bar charts. It also supports exporting in Excel format and API calls, with the API interface used by the power grid dispatch center.

[0034] The following steps verify the results. Step S1: Multi-source data acquisition Data collected: The total amount of wind turbine operation data is approximately 1.8144 × 10⁻⁶. 8 The calculation method is 10Hz × 30 units × 7 days; the total amount of wind measurement data is approximately 3.6288 × 10 5 The data is calculated as 1Hz × 6 stations × 7 days; there are 672 short-term meteorological data, 168 medium-term data, and 7 long-term data; there are 5 extreme weather warning messages, the content of which is a blue warning for Typhoon Talim. Data integrity: The integrity of the original data is ≥99.5%.

[0035] Step S2: Intelligent Data Preprocessing Abnormal data removal: A total of 286 abnormal data were detected, with a removal rate of approximately 0.00016%. The main abnormal data were three wind speed jumps caused by vibration of the west wind measurement tower during the typhoon. Missing data repair: The improved WGAN-GP was used to complete the missing data of the East Anemometer Tower. After completion, the wind speed data range is 10.2-13.5 m / s, and the deviation from the data of the surrounding anemometer towers during the same period is ≤3%. The wind speed data of the surrounding anemometer towers during the same period range is 10.5-13.8 m / s. Standardization results: Numerical features such as wind speed and temperature are normalized to the range of [-1,1], while categorical features such as equipment operating status are converted into one-hot encoded vectors. The output is a standardized dataset, in which the short-term dataset has 15 dimensions, the medium-term dataset has 12 dimensions, and the long-term dataset has 8 dimensions.

[0036] Step S3: Multi-scale Interaction Modeling and Prediction Initial prediction results: Short-term sub-model: Initial predicted power is 13.2MW, 12.8MW, 12.5MW, and 12.1MW respectively; Intermediate sub-model: The initial predicted average power for the entire day is 11.8 MW; Long-term sub-model: The initial forecast for daily power generation is 283 MWh.

[0037] Multi-scale interactive correction: Long-term submodel output trend coefficient The "daily average wind speed feature weight" of the mid-term sub-model was adjusted from 0.3 to 0.36, and the mid-term predicted average power was corrected to... ; Mid-term sub-model output load fluctuation coefficient The "grid load characteristic input value" of the short-term sub-model was revised from 0.65 to... The short-term power forecasts, after correction, are 13.5MW, 13.1MW, 12.8MW, and 12.4MW, respectively. The measured power of the short-term sub-model from 14:00 to 14:15 was 13.3MW, and the calculation deviation was... The corresponding error rate is 1.5%, which does not trigger the long-term sub-model parameter update, and the long-term daily power generation forecast remains at 283 MWh; Revised results: Short-term forecast results Mid-term forecast results Long-term forecast results .

[0038] Step S4: Dynamic Feature Weighting Scene recognition: A blue warning for Typhoon Talim was received. The maximum wind speed at the center of the typhoon was 12 m / s, and the distance to the wind farm was 180 km. The scene was switched from "normal weather" to "typhoon scene". Weighting adjustment: Feature importance scores in the "Typhoon Scenario": Wind direction stability 0.4, Typhoon path distance 0.2, Equipment wind resistance threshold 0.3, Other features 0.1; Calculation weights: wind direction stability weight 0.38, typhoon path distance weight 0.19, equipment wind resistance threshold weight 0.28, other feature weights 0.15; Updated forecasts: The short-term sub-model's predicted power has been adjusted from 13.5MW to 13.3MW due to the increased weighting of wind direction stability; the medium-term sub-model's predicted average power has been adjusted from 14.16MW to 13.8MW due to the addition of typhoon path distance weighting; and the long-term sub-model's predicted daily power generation has been adjusted from 283MWh to 278MWh due to the increased weighting of equipment wind resistance threshold. Weighted results: Short-term forecast results Mid-term forecast results Long-term forecast results .

[0039] Step S5: Real-time calibration Calculation of rolling window deviation: First time window: predicted average power 13.3MW, measured average power 13.2MW, calculation deviation. The corresponding error rate is 0.75%. Second time window: predicted average power 13.1MW, measured average power 13.0MW, calculation deviation. The corresponding error rate is 0.77%. Model fine-tuning: Mini-batchSGD fine-tuning was not triggered if the deviation of two consecutive windows was ≤1%. Prediction Correction: The calculation method for the next time-window prediction power is as follows ; Corrected results: Short-term forecast results Mid-term forecast results Long-term forecast results .

[0040] Step S6: Prediction Result Fusion and Output Confidence calculation: The prediction errors of the short-term sub-model in the last 10 predictions are 0.2MW, 0.15MW, 0.1MW, 0.12MW, 0.08MW, 0.1MW, 0.15MW, 0.09MW, 0.11MW, and 0.1MW, respectively. The mean error is calculated. The average of the last 10 measurements Confidence level ; The prediction errors of the mid-term sub-model in the last 10 predictions were 0.3MW, 0.25MW, 0.2MW, 0.18MW, 0.22MW, 0.25MW, 0.3MW, 0.28MW, 0.2MW, and 0.22MW, respectively. The mean error was calculated. The average of the last 10 measurements Confidence level ; The prediction errors of the long-term sub-model in the last 10 predictions were 5 MWh, 4.5 MWh, 3.8 MWh, 4.2 MWh, 3.5 MWh, 4.0 MWh, 4.5 MWh, 3.9 MWh, 4.1 MWh, and 3.7 MWh, respectively. The mean error was calculated. The average of the last 10 measurements Confidence level ; Short-term final power: Calculated as follows The result is that 278 MWh divided by 24 hours equals approximately 11.58 MW. ; Medium-term final average power: The calculation method is the same as for the short-term, and the final result is approximately... ; Long-term final daily power generation: Calculation method is as follows final result ; Uncertainty range: Short-term error standard deviation The uncertainty interval is ; Long-term error standard deviation The uncertainty interval is ; Output results: The Web platform outputs three core results: short-term power 12.6MW, medium-term average power 12.6MW, and daily power generation 272MWh.

[0041] In summary, this invention significantly improves the accuracy, real-time performance, and robustness of wind farm power generation and output prediction through multi-source data fusion, multi-scale interactive modeling, dynamic feature weighting, and real-time correction, effectively supporting grid dispatch optimization and safe operation of wind farms.

[0042] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it.

Claims

1. A wind farm power generation and electricity generation prediction system, characterized in that, include: The multi-source data acquisition module is used to collect wind turbine operation data, on-site wind measurement data, short-term / medium-term / long-term meteorological data, extreme weather warning data, power grid load data, and wind farm environmental data. The intelligent data preprocessing module is used to perform anomaly detection and removal, missing value repair, and standardization processing on the collected data; The multi-scale interactive modeling module includes a short-term sub-model, a medium-term sub-model, a long-term sub-model, and a multi-scale interactive unit. The short-term sub-model adopts Bi-LSTM, the medium-term sub-model adopts XGBoost, the long-term sub-model adopts an improved Transformer, and the multi-scale interactive unit is used to realize parameter transfer and result correction between sub-models. The dynamic feature weighting module is used to adjust feature weights in real time based on scene recognition results through an attention mechanism; The real-time correction module is used to calculate the prediction bias using a rolling window mechanism, and to fine-tune the model parameters and correct the prediction results based on the bias. The prediction result fusion and output module is used to fuse the results of multiple sub-models based on confidence weights and output the predicted values ​​and uncertainty intervals.

2. The wind farm power generation and electricity generation prediction system according to claim 1, characterized in that, The real-time sensor data sampling frequency of the multi-source data acquisition module is 10Hz, and the meteorological data update frequency is: 15 minutes / time for short-term, 1 hour / time for medium-term, and 1 day / time for long-term. The multi-source data acquisition module also includes an extreme weather warning API interface.

3. The wind farm power generation and electricity generation prediction system according to claim 2, characterized in that: The intelligent data preprocessing module uses the isolated forest algorithm to detect abnormal data, the improved WGAN-GP to complete missing data, Z-score normalization to process numerical features, and one-hot encoding to process categorical features.

4. The wind farm power generation and electricity generation prediction system according to claim 3, characterized in that: The operation of the multi-scale interaction unit includes: adjusting the feature weights of the intermediate sub-model with the trend coefficient output by the long-term sub-model; correcting the input features of the short-term sub-model with the volatility coefficient output by the intermediate sub-model; and adjusting the trend prediction of the long-term sub-model with the deviation feedback output by the short-term sub-model.

5. The wind farm power generation and electricity generation prediction system according to claim 3, characterized in that: The dynamic feature weighting module includes a scene recognition unit, an attention weight calculation unit, and a weight update unit. The scene recognition unit can recognize "normal weather", "typhoon warning" and "cold wave" scenes. The attention weight calculation unit uses the softmax function to calculate feature weights. The weight update unit updates the weights every 30 minutes.

6. The wind farm power generation and electricity generation prediction system according to claim 5, characterized in that: The real-time correction module has a rolling window duration of 1 hour, uses Mini-batchSGD to fine-tune the sub-model output layer parameters, and inversely superimposes the deviation values ​​into the prediction results of the next window.

7. The wind farm power generation and electricity generation prediction system according to claim 6, characterized in that: The confidence level of the prediction result fusion and output module is calculated based on the prediction error of the sub-model in the last 10 predictions, and the uncertainty range is "predicted value ± 2 × standard deviation of the error in the last 10 predictions".

8. A method for predicting the power generation and electricity output of a wind farm, characterized in that, Includes the following steps: S1: Collect multi-dimensional data through a multi-source data acquisition module; S2: The intelligent data preprocessing module performs anomaly removal, missing data repair, and standardization. S3: Through the multi-scale interactive modeling module, the preprocessed data is input into the short-term, medium-term and long-term sub-models, and the prediction results are corrected through the multi-scale interactive unit to obtain P1, P2 and P3; S4: Through the dynamic feature weighting module, adjust the feature weights based on scene recognition and update P1, P2, and P3; S5: Through the real-time correction module, the deviation is calculated using a rolling window, the model parameters are fine-tuned and the prediction results are corrected to obtain P1', P2', and P3'. S6: Through the prediction result fusion and output module, P1', P2', and P3' are fused based on confidence weights to output the final predicted value and uncertainty range.

9. The method for predicting wind farm power generation and electricity output according to claim 8, characterized in that: The short-term sub-model is Bi-LSTM, which takes data from the next 2 hours as input; the medium-term sub-model is XGBoost, which takes data from the next 7 days as input; and the long-term sub-model is an improved Transformer, which takes data from the next 1 year as input.

10. The method for predicting wind farm power generation and electricity output according to claim 8, characterized in that: In step S6, the formula for calculating the final predicted value P is: C1, C2, and C3 represent the confidence levels of the short-term, medium-term, and long-term sub-models, respectively, and the confidence level is the average of the prediction errors from 1 to the last 10 predictions.

Images