Wind power generation power prediction method and system based on big data
By using a big data-based wind power generation prediction method, and employing causal discovery algorithms and physical boundary verification, the problem of inaccurate power generation prediction caused by the subtle influence of meteorological data on the unit's operating status is solved, thus achieving more accurate power generation prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUANENG (TIANJIN) CLEAN ENERGY CO LTD
- Filing Date
- 2026-02-26
- Publication Date
- 2026-06-05
AI Technical Summary
In actual power generation, different meteorological data can have a slight impact on the unit's operating status, leading to deviations in operating parameters and making it impossible to accurately predict power generation.
The big data-based wind power generation prediction method acquires historical multi-source data, performs timestamp alignment and time-frequency decomposition, extracts multi-scale feature components, uses a causal discovery algorithm to analyze the causal relationship between meteorological data, unit operating parameters and power generation, constructs and trains a power generation prediction model, uses the meteorological prediction model to predict future meteorological sequences and perform causal relationship path propagation inference, and combines physical boundary constraint verification and time integration to calculate the predicted power generation.
It improves the accuracy of power generation forecasting, avoids forecasting errors caused by signal interference, clarifies the causal relationship between meteorological data and unit operating parameters, and can deduce subtle impacts on the unit's operating status from future meteorological data, correct abnormal forecast values, and ensure that forecast values are within the unit's rated power range.
Smart Images
Figure CN122159188A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of wind power generation prediction, and in particular to a method and system for predicting wind power generation based on big data. Background Technology
[0002] As a major form of renewable energy, wind power has brought great convenience to human production and life. In the field of wind power generation, accurate prediction of power generation is extremely important. There are already many prediction methods in existing technologies, such as relying on large-scale weather forecasts from meteorological stations and relying on historical data of wind turbines for prediction.
[0003] However, in actual power generation, various weather conditions are encountered, such as heavy rain and turbulence. These adverse weather conditions can affect the operating status of the turbines. For example, turbulence has a subtle impact on the turbine's rotational speed, causing deviations in the actual turbine speed and making it impossible to accurately predict power generation. Based on this, the applicant proposes a wind power generation prediction method and system based on big data. Summary of the Invention
[0004] The technical problem solved by this invention is that different meteorological data during actual power generation can have a slight impact on the operating status of the unit, resulting in deviations in operating parameters and making it impossible to accurately predict power generation.
[0005] To address the aforementioned technical problems, the first aspect of this invention provides the following technical solution: a wind power generation prediction method based on big data, comprising: Step S1: Obtain historical multi-source data, align the historical multi-source data with timestamps, perform time-frequency decomposition, and extract multi-scale feature components; Step S2: Based on the multi-scale feature components, a causal discovery algorithm is used to analyze the causal relationship path between meteorological data, unit operating parameters and power generation, forming a causal relationship network. Based on the causal relationship network, key variables are selected, and a power generation prediction model is constructed and trained. Step S3: Predict future weather sequences using a meteorological forecasting model, perform causal path propagation deduction based on the future weather sequences, generate a deduced sequence of unit operating parameters, and then input the future weather sequences and the deduced sequence of unit operating parameters into the power generation prediction model to obtain the predicted power generation. Step S4: Perform physical boundary constraint verification on the predicted power generation based on the upper and lower limits of the unit power curve corresponding to the current wind speed, correct abnormal prediction values, and obtain the final predicted power generation. Step S5: Based on the final predicted power generation, calculate the predicted power generation for future periods through time integration.
[0006] Preferably, in step S1, the historical multi-source data includes historical meteorological data, historical unit operating parameters, and historical power generation. The historical meteorological data includes wind speed, wind direction, air pressure, temperature, and turbulence intensity. The historical unit operating parameters include generator speed, blade pitch angle, bearing temperature, gearbox vibration, and yaw angle. The historical power generation includes power generation at the same time scale as the historical meteorological data and historical unit operating parameters.
[0007] Preferably, step S1 specifically includes: Step S11: Clean the historical multi-source data, process missing values and outliers, and align them with ten-minute precision based on a unified timestamp to form a synchronized multi-source time series. Step S12: Determine the sliding window length T and the sliding step size, divide the multivariate time series into multiple samples, and for each sample, use the meteorological and unit operation parameters of the past T time points as input variables and the power generation of the next time point as the label; Step S13: Perform EMD decomposition on the input variables of each sample in the multivariate time series to obtain the first IMF component and the first residual. Combine the first K IMF components of all input variables into a three-dimensional tensor, wherein the three dimensions include the scale dimension, the time dimension and the variable dimension.
[0008] Preferably, step S2 specifically includes: Step S21: Perform the same EMD decomposition on the label of each sample in step S12 as on the input variable to obtain the second IMF component and the second residual of the power generation label; Step S22: Fuse the second IMF component of the three-dimensional tensor to generate a four-dimensional tensor, wherein the four dimensions include scale dimension, time dimension, variable type dimension and specific variable dimension; Step S23: Input the four-dimensional tensor into the PCMCI+ algorithm to obtain a directed acyclic graph; Step S24: Based on causal strength and knowledge of wind power generation, calculate the cumulative causal contribution of each original variable to the power component across all IMF scales, and screen out key original variables whose causal influence exceeds a preset threshold. Step S25: Perform feature extraction on the key original variables, specifically including: By using a CNN network to perform convolutions along the variable dimension, the time series of different key original variables at the same time are extracted, and the final feature vector is obtained. Step S26: Input the final feature vector into the GRU for training to obtain the power generation prediction model.
[0009] Preferably, step S2 further includes: Step S27: Construct a set of meteorological-unit parameter influence equations based on the causal relationship network; Step S271: In the causal relationship network, extract all causal edges from the meteorological variable IMF component to the key unit operating parameter IMF component; Step S272: Construct a set of structural equations based on each causal edge, and estimate the parameters in each equation using the maximum likelihood method; Step S273: Perform residual analysis on each structural equation to ensure that the error of each equation is within the first allowable threshold; Step S274: Integrate all the constructed structural equations to obtain the meteorological-unit parameter influence equation set.
[0010] Preferably, step S3 specifically includes: Step S31: Call the ENS of ECMWF to obtain the future meteorological sequence for the future target time period, perform EMD decomposition on the future meteorological sequence, and obtain the third IMF component and the third residual for each future meteorological variable; Step S32: Take the IMF component values of the future meteorological sequence at time step t as exogenous input, take the IMF component values of the unit operating parameters at time step t-1 as endogenous input, substitute them into the meteorological-unit parameter influence equation set, solve the equation set, and calculate the extrapolated values of each IMF component of the unit operating parameters at time step t. Step S33: Store the extrapolated values of each IMF component of the unit operating parameters at time step t into the corresponding future sequence; Step S34: Move to time step t+1 and repeat steps S32-S33 until the extrapolation of all future time steps is completed, and obtain the extrapolation sequence of IMF components of unit operating parameters. Step S35: Feature extraction based on the inference sequence of the third IMF component of the future meteorological variable and the IMF component of the unit operating parameters specifically includes: using CNN to convolve in the time dimension to extract the time series of the inference sequence of the third IMF component of the future meteorological variable and the IMF component of the unit operating parameters at the same time, to obtain the second final feature vector, inputting the second final feature vector into the power generation prediction model according to the time step, and taking the hidden state of the last time step as the feature vector of the future period.
[0011] Preferably, step S4 specifically includes: Step S41: Obtain the predicted power generation value and obtain the wind speed at the time point corresponding to the predicted power generation value; Step S42: Obtain the rated power curve of the unit at the wind speed from the data of the unit manufacturer, and expand the rated power curve upward and downward by a preset percentage to form an operating envelope, and obtain the upper limit and lower limit of the envelope; Step S43: Obtain the predicted power generation value for this time step, and determine whether it is within the upper and lower limits of the envelope. If it is not within the envelope, it is considered an abnormal prediction value and is truncated and corrected. If it is within the envelope, it is considered a normal prediction value. Integrate the normal prediction value and the corrected abnormal prediction value into the final predicted power generation value.
[0012] Preferably, in step S43, truncating and correcting the abnormal predicted values includes: If the predicted power generation is greater than the upper limit of the envelope, the predicted power generation will be corrected to the upper limit of the envelope. If the predicted power generation is less than the lower envelope limit, the predicted power generation will be corrected to the lower envelope limit value.
[0013] Preferably, step S5 specifically includes: The total power generation E during the future target period is calculated using numerical integration methods such as the trapezoidal rule. ; Where E represents the total power generation within the target future period. This represents the number of discrete points in the target time period. Indicates the time step. Let represent the final predicted power generation at time point i. This represents the final predicted power generation at time point i+1.
[0014] Secondly, a wind power generation prediction system based on big data is provided to execute the wind power generation prediction method based on big data, including: a data acquisition and processing module, a model building module, a power generation prediction module, a verification and correction module, and a power generation calculation module. The data acquisition and processing module is used to acquire historical multi-source data, align the historical multi-source data with timestamps, perform time-frequency decomposition, and extract multi-scale feature components. The model building module is used to analyze the causal relationship path between meteorological data, unit operating parameters and power generation based on the multi-scale feature components and the causal discovery algorithm, form a causal relationship network, screen key variables based on the causal relationship network, and build and train a power generation prediction model. The power generation prediction module is used to predict future weather sequences using a meteorological prediction model, perform causal path propagation deduction based on the future weather sequences, generate a deduction sequence of unit operating parameters, and then input the future weather sequences and the deduction sequence of unit operating parameters into the power generation prediction model to obtain the predicted power generation. The verification and correction module is used to verify the physical boundary constraints of the predicted power generation based on the upper and lower limits of the unit power curve corresponding to the current wind speed, correct abnormal prediction values, and obtain the final predicted power generation. The power generation calculation module is used to calculate the predicted power generation for future periods based on the final predicted power generation through time integration.
[0015] The beneficial effects of this invention are as follows: This invention performs time-frequency decomposition on the original signal, enabling targeted learning of features at different scales and avoiding prediction errors caused by signal mixing; through a causal discovery algorithm, it clearly reveals the causal relationship between meteorological data, unit operating parameters, and power generation, tracing specific causal links, enabling the deduction of subtle influences of meteorological conditions on the unit's operating state from future meteorological data, and quantitative calculations through structural equation modeling to obtain the actual operating parameters of the unit, and then power generation prediction based on meteorological data and actual operating parameters of the unit, thereby improving the accuracy of power generation prediction. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the basic process of a wind power generation prediction method based on big data, provided as an embodiment of the present invention. Detailed Implementation
[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0018] Example 1, referring to Figure 1 As one embodiment of the present invention, a wind power generation prediction method based on big data is provided, comprising: Step S1: Obtain historical multi-source data, align the historical multi-source data with timestamps, perform time-frequency decomposition, and extract multi-scale feature components; Step S2: Based on the multi-scale feature components, the causal discovery algorithm is used to analyze the causal relationship path between meteorological data, unit operating parameters and power generation, form a causal relationship network, select key variables based on the causal relationship network, and build and train the power generation prediction model. Step S3: Use a meteorological forecasting model to predict future weather sequences, perform causal path propagation deduction based on the future weather sequences, generate a deduced sequence of unit operating parameters, and then input the future weather sequences and the deduced sequence of unit operating parameters into the power generation prediction model to obtain the predicted power generation. Step S4: Perform physical boundary constraint verification on the predicted power generation based on the upper and lower limits of the unit power curve corresponding to the current wind speed, correct abnormal prediction values, and obtain the final predicted power generation. Step S5: Based on the final predicted power generation, calculate the predicted power generation for future periods through time integration.
[0019] This invention performs time-frequency decomposition on historical multi-source data, enabling targeted learning of features at different scales and avoiding prediction errors caused by signal confounding. Through a causal discovery algorithm, it clearly reveals the causal relationships between meteorological data, unit operating parameters, and power generation, tracing specific causal links. This allows for the deduction of subtle influences of weather on unit operating conditions from future meteorological data, obtaining the actual operating parameters of the unit. Power generation is then predicted based on the meteorological data and actual operating parameters, improving the accuracy of power generation forecasts. After power generation prediction is completed, abnormal prediction values are corrected to prevent them from falling outside the unit's rated power range.
[0020] In step S1, the historical multi-source data includes historical meteorological data, historical unit operating parameters, and historical power generation. Historical meteorological data includes wind speed, wind direction, air pressure, temperature, and turbulence intensity; historical unit operating parameters include generator speed, blade pitch angle, bearing temperature, gearbox vibration, and yaw angle; historical power generation includes power generation at the same time scale as historical meteorological data and historical unit operating parameters. Step S1 specifically includes: Step S11: Clean the historical multi-source data, handle missing values and outliers, and align them with ten-minute precision based on a unified timestamp to form a synchronized multivariate time series. In this embodiment, for missing values at a certain moment, the average of the two valid values before and after is used to fill in the missing values. Obviously unreasonable outliers are deleted; for example, a wind speed sensor malfunction causing a wind speed significantly higher than normal is considered an outlier. The average of all wind turbine speed data from 9:00 to 9:10 is calculated at 10-minute intervals to obtain the speed data for that 10-minute timestamp (9:10). After aligning all processed historical multi-source data, a synchronized multivariate time series is formed. Rows represent a 10-minute timestamp, and columns represent wind speed (m / s), wind direction (°), air pressure (hPa), temperature (°C), turbulence intensity, generator speed (rpm), blade pitch angle (°), bearing temperature (°C), gearbox vibration (mm / s), yaw angle (°), and power generation.
[0021] Step S12: Determine the sliding window length T and the sliding step size, divide the multivariate time series into multiple samples, and for each sample, use the meteorological and unit operation parameters of the past T time points as input variables and the power generation of the next time point as the label; In this embodiment, the sliding window length T is set to 6, with one hour as one time window, and each time window has 6 ten-minute intervals. The sliding step size is set to 10 minutes. The multivariate time series is divided chronologically: the first 70% is the training set, the middle 15% is the validation set, and the last 15% is the test set.
[0022] Step S13: Perform EMD decomposition on the input variables of each sample in the multivariate time series to obtain the first IMF component and the first residual. Combine the first K IMF components of all input variables into a three-dimensional tensor, where the three dimensions include the scale dimension, the time dimension and the variable dimension. This embodiment uses Empirical Mode Decomposition (EMD) to decompose non-stationary and nonlinear meteorological data and unit operating parameters into intrinsic mode functions (IMFs) at different time scales, providing a foundation for multi-scale causal analysis.
[0023] For variables Decompose using the following formula: ; in, This represents the value of the variable x to be decomposed at time point t. Indicates the kth The value of the component at time point t, where K represents the number of IMF components retained. This represents the residual term.
[0024] Extract the first K Intra-Functional Components (IMFs) from all meteorological data and unit operating parameters (M in total). Each IMF is a time series of length T. Stack the IMFs of all variables into a three-dimensional array. The scale dimension includes high-frequency, mid-frequency, and low-frequency data; the time dimension has T historical time points; and the variable dimension has M original variables. For example, if the scale dimension is 3, T is 6, there are 5 meteorological data points and 5 unit operating parameters, then M is 10, resulting in a tensor shape of (3, 6, 10). Step S1 decomposes the mixed raw data into different scales to avoid feature confusion. After decomposition, the high-frequency IMF can be regarded as a rapidly varying perturbation, which is convenient for subsequent analysis of subtle effects. This step provides input for causal discovery in the subsequent step S2, enabling the PCMCI+ algorithm to identify cross-scale causal paths such as "high-frequency turbulence → mid-frequency speed fluctuation → low-frequency power change".
[0025] Step S2 specifically includes: Step S21: Perform the same EMD decomposition on the label of each sample in step S12 as on the input variables to obtain the second IMF component and the second residual of the power generation label; In this embodiment, an EMD decomposition is performed on a historical power generation of 1500kW. After decomposition, the minute-level fluctuation caused by turbulence is obtained as +20kW, and the ten-minute-level fluctuation caused by gusts is obtained as +50kW.
[0026] This step performs the same EMD decomposition on the power generation labels as on the input variables, placing them in the same multi-scale representation space as the input features.
[0027] Step S22: Fuse the second IMF component of the three-dimensional tensor to generate a four-dimensional tensor, which includes the scale dimension, time dimension, variable type dimension, and specific variable dimension. In this embodiment, K is set to 3 and T is set to 6. At the high frequency scale, there are 5 meteorological data and 5 unit parameter data at time points 1-6, and 1 power data at time point 7. At the mid-frequency and low-frequency scales, there are data with the same power as at the high frequency scale. The final shape is (3, 7, 3, max(5, 5, 1)).
[0028] Step S23: Input the four-dimensional tensor into the PCMCI+ algorithm to obtain a directed acyclic graph; The PCMCI+ algorithm is run independently at each scale, and the causal graph at each scale includes nodes, directed edges, and causal strength. The causal graphs at all scales are then integrated to label cross-scale causal relationships, forming a complete multi-scale causal network of "meteorology-unit parameters-power".
[0029] Step S24: Based on causal strength and knowledge of wind power generation, calculate the cumulative causal contribution of each original variable to the power component across all IMF scales, and screen out key original variables whose causal influence exceeds a preset threshold. For each original variable X, its cumulative causal contribution across scales is calculated using the following formula: ; in, Represents the original variable The cumulative causal contribution, where K represents the number of IMF components. This represents the IMF component of the original variable at the i-th scale. This represents the IMF component of power generation at the j-th scale. Indicates from arrive Causal strength weights.
[0030] Step S25: Feature extraction of key original variables, specifically including: By using a CNN network to perform convolutions along the variable dimension, the time series of different key original variables at the same time are extracted, and the final feature vector is obtained. Step S26: Input the final feature vector into the GRU for training to obtain the power generation prediction model.
[0031] Step S2 also includes: Step S27: Construct a set of equations relating meteorological and generator parameters based on a causal relationship network; Step S271: In the causal relationship network, extract all causal edges from the meteorological variable IMF component to the key unit operating parameter IMF component; Step S272: Construct a system of structural equations for each causal edge, and estimate the parameters in each equation using the maximum likelihood method; Step S273: Perform residual analysis on each structural equation to ensure that the error of each equation is within the first allowable threshold; Step S274: Integrate all the constructed structural equations to obtain the meteorological-unit parameter influence equation set.
[0032] In a causal network, extract the causal edges from meteorological variables to unit variables. At the same time, it is necessary to pay attention to the scale of these variables, such as high-frequency wind speed → medium-frequency turbine speed, medium-frequency wind speed → medium-frequency blade pitch angle, high-frequency turbulence → medium-frequency turbine speed, etc.
[0033] Based on the extracted causal edges, a structural equation is constructed for each affected unit parameter. The right side of the equation represents all meteorological factors affecting the unit parameter and the value of the unit parameter at the previous time step. The parameters in the equation are estimated using the maximum likelihood method.
[0034] In this embodiment, a structural equation relating wind speed and turbulence intensity was constructed based on a causal relationship network as follows: Fan speed (t) = 0.75 × wind speed (t) + 0.1 × turbulence intensity (t) + 0.85 × fan speed (t-1) + 0.5; The above structural equations indicate that for every unit increase in wind speed, the fan speed increases by 0.75 units; for every unit increase in turbulence, the fan speed increases by approximately 0.1 units; and 0.85 represents the fan speed, indicating that it has strong inertia.
[0035] In step S273, the fitted equation is used to perform another calculation on the historical data. The calculated values are subtracted from the actual values to obtain a series of differences, which are used as residuals. The closer the residuals are to 0, the stronger the reliability of the equation. Therefore, an acceptable fluctuation range needs to be set. If the mean absolute residuals are within this acceptable fluctuation range, the equation is considered reliable.
[0036] Step S3 specifically includes: Step S31: Call the ENS of ECMWF to obtain the future meteorological sequence for the future target time period, perform EMD decomposition on the future meteorological sequence, and obtain the third IMF component and third residual for each future meteorological variable; ECMWF's ENS is a relatively mature existing weather forecasting system. It obtains the future weather sequence for the target time period and performs EMD decomposition on it. For example, it decomposes wind speed data into high-frequency IMF components, mid-frequency IMF components, and low-frequency IMF components.
[0037] Step S32: Take the IMF component values of the future meteorological sequence at time step t as exogenous inputs, and take the IMF component values of the unit operating parameters at time step t-1 as endogenous inputs. Substitute them into the meteorological-unit parameter influence equation set, solve the equation set, and calculate the extrapolated values of each IMF component of the unit operating parameters at time step t. In this embodiment, the formula is: wind turbine speed intermediate frequency component (t) = 0.75 × wind speed high frequency component (t) + 0.1 × turbulence intensity high frequency component (t) + 0.85 × wind turbine speed intermediate frequency component (t-1) + 0.5; Substituting the high-frequency components of wind speed and turbulence at the current time step, as well as the mid-frequency components of wind turbine speed at the previous time step, into the structural equation, we obtain the mid-frequency component of wind turbine speed at the current time step, which is the derived value.
[0038] For example, at 2:00 PM, the weather forecast indicates that the high-frequency component of wind speed will increase from 8 to 8.5 by 2:10 PM, the high-frequency component of turbulence is 0.12, and the mid-frequency component of the wind turbine speed at 2:00 PM is 12. Substituting these values into the formula, the projected value of the mid-frequency component of the wind turbine speed at 2:10 PM is: The intermediate frequency component of the fan speed (14:10) = 0.75 × 8.5 + 0.1 × 0.12 + 0.85 × 12 + 0.5 = 17.1 Therefore, the mid-frequency component of the rotational speed at 14:10 is approximately 17.1.
[0039] Step S33: Store the extrapolated values of each IMF component of the unit operating parameters at time step t into the corresponding future sequence; Step S34: Move to time step t+1 and repeat steps S32-S33 until the extrapolation of all future time steps is completed, and obtain the extrapolation sequence of IMF components of unit operating parameters. Step S35: Feature extraction based on the inference sequence of the third IMF component of future meteorological variables and the IMF component of unit operating parameters specifically includes: using CNN to convolve in the time dimension to extract the time series of the inference sequence of the third IMF component of future meteorological variables and the IMF component of unit operating parameters at the same time, to obtain the second final feature vector, inputting the second final feature vector into the power generation prediction model according to the time step, and taking the hidden state of the last time step as the feature vector of the future period.
[0040] Step S4 specifically includes: Step S41: Obtain the predicted power generation value and the wind speed at the time point corresponding to the predicted power generation value; Step S42: Obtain the rated power curve of the unit at the wind speed from the data of the unit manufacturer, expand the rated power curve upward and downward by a preset percentage to form an operating envelope, and obtain the upper and lower limits of the envelope. Step S43: Obtain the predicted power generation value for this time step, and determine whether it is within the upper and lower limits of the envelope. If it is not within the envelope, it is considered an abnormal prediction value and is truncated and corrected. If it is within the envelope, it is considered a normal prediction value. Integrate the normal prediction value and the corrected abnormal prediction value into the final predicted power generation value.
[0041] During the forecasting process, predicted values may appear that significantly exceed the equipment's capabilities. For example, when the wind speed is 15, the rated power curve of the unit shows that the rated power must be between 9kW and 10kW. Considering that this curve is an ideal value and there is an error, the preset percentage is extended upward and downward to add a safety buffer zone to the theoretical range.
[0042] If the calculated predicted power generation value exceeds the boundary of the envelope, it is judged as an abnormal predicted value; if it does not exceed the boundary, it is judged as a normal predicted value.
[0043] If the predicted power generation is greater than the upper limit of the envelope, the predicted power generation will be corrected to the upper limit of the envelope. If the predicted power generation is less than the lower envelope limit, the predicted power generation will be corrected to the lower envelope limit value.
[0044] Step S5 specifically includes: calculating the total power generation E within the future target time period using numerical integration methods such as the trapezoidal rule. ; Where E represents the total power generation within the target future period. This represents the number of discrete points in the target time period. Indicates the time step. Let represent the final predicted power generation at time point i. This represents the final predicted power generation at time point i+1.
[0045] In this embodiment, the time step is 10 minutes, and each 10 minutes is considered as a time period. Only one power value is obtained in each time period. The final predicted power generation value is discrete. Therefore, numerical integration methods such as the trapezoidal rule are used to calculate the total power generation.
[0046] Example 2 provides a wind power generation prediction system based on big data. The wind power generation prediction system based on big data includes: a data acquisition and processing module, a model building module, a power generation prediction module, a verification and correction module, and a power generation calculation module. The data acquisition and processing module is used to acquire historical multi-source data, align the historical multi-source data with timestamps, perform time-frequency decomposition, and extract multi-scale feature components. The model building module is used to analyze the causal relationship paths between meteorological data, unit operating parameters and power generation based on multi-scale feature components and causal discovery algorithms, form a causal relationship network, screen key variables based on the causal relationship network, and build and train a power generation prediction model. The power generation prediction module is used to predict future weather sequences using a meteorological prediction model, perform causal path propagation deduction based on the future weather sequences, generate a deduced sequence of unit operating parameters, and then input the future weather sequences and the deduced sequence of unit operating parameters into the power generation prediction model to obtain the predicted power generation. The verification and correction module is used to verify the physical boundary constraints of the predicted power generation based on the upper and lower limits of the unit power curve corresponding to the current wind speed, correct abnormal prediction values, and obtain the final predicted power generation. The power generation calculation module is used to calculate the predicted power generation for future periods based on the final predicted power generation through time integration.
[0047] This invention performs time-frequency decomposition on the original signal, enabling targeted learning of features at different scales and avoiding prediction errors caused by signal confounding. Through a causal discovery algorithm, it clearly reveals the causal relationship between meteorological data, unit operating parameters, and power generation, tracing specific causal links. This allows for the deduction of subtle influences of weather on unit operating status from future meteorological data. Quantitative calculations using structural equation modeling yield the actual operating parameters of the unit. Power generation prediction is then made based on meteorological data and actual operating parameters, improving the accuracy of power generation prediction.
[0048] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0049] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.
Claims
1. A wind power generation prediction method based on big data, characterized in that, include: Step S1: Obtain historical multi-source data, align the historical multi-source data with timestamps, perform time-frequency decomposition, and extract multi-scale feature components; Step S2: Based on the multi-scale feature components, a causal discovery algorithm is used to analyze the causal relationship path between meteorological data, unit operating parameters and power generation, forming a causal relationship network. Based on the causal relationship network, key variables are selected, and a power generation prediction model is constructed and trained. Step S3: Predict future weather sequences using a meteorological forecasting model, perform causal path propagation deduction based on the future weather sequences, generate a deduced sequence of unit operating parameters, and then input the future weather sequences and the deduced sequence of unit operating parameters into the power generation prediction model to obtain the predicted power generation. Step S4: Perform physical boundary constraint verification on the predicted power generation based on the upper and lower limits of the unit power curve corresponding to the current wind speed, correct abnormal prediction values, and obtain the final predicted power generation. Step S5: Based on the final predicted power generation, calculate the predicted power generation for future periods through time integration.
2. The wind power generation prediction method based on big data as described in claim 1, characterized in that: In step S1, the historical multi-source data includes historical meteorological data, historical unit operating parameters, and historical power generation. The historical meteorological data includes wind speed, wind direction, air pressure, temperature, and turbulence intensity. The historical unit operating parameters include generator speed, blade pitch angle, bearing temperature, gearbox vibration, and yaw angle. The historical power generation includes power generation at the same time scale as the historical meteorological data and historical unit operating parameters.
3. The wind power generation prediction method based on big data as described in claim 2, characterized in that: Step S1 specifically includes: Step S11: Clean the historical multi-source data, process missing values and outliers, and align them with ten-minute precision based on a unified timestamp to form a synchronized multi-source time series. Step S12: Determine the sliding window length T and the sliding step size, divide the multivariate time series into multiple samples, and for each sample, use the meteorological and unit operation parameters of the past T time points as input variables and the power generation of the next time point as the label; Step S13: Perform EMD decomposition on the input variables of each sample in the multivariate time series to obtain the first IMF component and the first residual. Combine the first K IMF components of all input variables into a three-dimensional tensor, wherein the three dimensions include the scale dimension, the time dimension and the variable dimension.
4. The wind power generation prediction method based on big data as described in claim 3, characterized in that: Step S2 specifically includes: Step S21: Perform EMD decomposition on the label of each sample in step S12, using the same method as the input variable, to obtain the second IMF component and the second residual of the power generation label; Step S22: Fuse the second IMF component of the three-dimensional tensor to generate a four-dimensional tensor, wherein the four dimensions include a scale dimension, a time dimension, a variable type dimension, and a specific variable dimension; Step S23: Input the four-dimensional tensor into the PCMCI+ algorithm to obtain a directed acyclic graph; Step S24: Based on causal strength and knowledge of wind power generation, calculate the cumulative causal contribution of each original variable to the power component across all IMF scales, and screen out key original variables whose causal influence exceeds a preset threshold. Step S25: Perform feature extraction on the key original variables, specifically including: By using a CNN network to perform convolutions along the variable dimension, the time series of different key original variables at the same time are extracted, and the final feature vector is obtained. Step S26: Input the final feature vector into the GRU for training to obtain the power generation prediction model.
5. The wind power generation prediction method based on big data as described in claim 4, characterized in that: Step S2 also includes: Step S27: Construct a set of meteorological-unit parameter influence equations based on the causal relationship network; Step S271: In the causal relationship network, extract all causal edges from the meteorological variable IMF component to the key unit operating parameter IMF component; Step S272: Construct a set of structural equations based on each causal edge, and estimate the parameters in each equation using the maximum likelihood method; Step S273: Perform residual analysis on each structural equation to ensure that the error of each equation is within the first allowable threshold; Step S274: Integrate all the constructed structural equations to obtain the meteorological-unit parameter influence equation set.
6. The wind power generation prediction method based on big data as described in claim 5, characterized in that: Step S3 specifically includes: Step S31: Call the ENS of ECMWF to obtain the future meteorological sequence for the future target time period, perform EMD decomposition on the future meteorological sequence, and obtain the third IMF component and the third residual for each future meteorological variable; Step S32: Take the IMF component values of the future meteorological sequence at time step t as exogenous input, take the IMF component values of the unit operating parameters at time step t-1 as endogenous input, substitute them into the meteorological-unit parameter influence equation set, solve the equation set, and calculate the extrapolated values of each IMF component of the unit operating parameters at time step t. Step S33: Store the extrapolated values of each IMF component of the unit operating parameters at time step t into the corresponding future sequence; Step S34: Move to time step t+1 and repeat steps S32-S33 until the extrapolation of all future time steps is completed, and obtain the extrapolation sequence of IMF components of unit operating parameters. Step S35: Based on the third IMF component of the future meteorological variables and the IMF component extrapolation sequence of the unit operating parameters, feature extraction is performed to obtain the feature vector of the future time period. The feature vector of the future time period is input into the power generation prediction model to obtain the predicted power generation value of each future time step. Feature extraction based on the inference sequence of the third IMF component of the future meteorological variable and the IMF component of the unit operating parameters specifically includes: using CNN to convolve in the time dimension to extract the time series of the inference sequence of the third IMF component of the future meteorological variable and the IMF component of the unit operating parameters at the same moment, to obtain the second final feature vector, inputting the second final feature vector into the power generation prediction model according to the time step, and taking the hidden state of the last time step as the feature vector of the future period.
7. The wind power generation prediction method based on big data as described in claim 6, characterized in that: Step S4 specifically includes: Step S41: Obtain the predicted power generation value and obtain the wind speed at the time point corresponding to the predicted power generation value; Step S42: Obtain the rated power curve of the unit at the wind speed from the data of the unit manufacturer, and expand the rated power curve upward and downward by a preset percentage to form an operating envelope, and obtain the upper limit and lower limit of the envelope; Step S43: Obtain the predicted power generation value for this time step, and determine whether it is within the upper and lower limits of the envelope. If it is not within the envelope, it is considered an abnormal prediction value and is truncated and corrected. If it is within the envelope, it is considered a normal prediction value. Integrate the normal prediction value and the corrected abnormal prediction value into the final predicted power generation value.
8. The wind power generation prediction method based on big data as described in claim 7, characterized in that: In step S43, Truncation correction for outlier predictions includes: If the predicted power generation is greater than the upper limit of the envelope, the predicted power generation will be corrected to the upper limit of the envelope. If the predicted power generation is less than the lower envelope limit, the predicted power generation will be corrected to the lower envelope limit value.
9. The wind power generation prediction method based on big data as described in claim 8, characterized in that: Step S5 specifically includes: calculating the total power generation E within the future target time period using numerical integration methods such as the trapezoidal rule. ； Where E represents the total power generation within the target future period. This represents the number of discrete points in the target time period. Indicates the time step. Let represent the final predicted power generation at time point i. This represents the final predicted power generation at time point i+1.
10. A wind power generation prediction system based on big data, characterized in that: The wind power generation prediction system based on big data includes: a data acquisition and processing module, a model building module, a power generation prediction module, a verification and correction module, and a power generation calculation module. The data acquisition and processing module is used to acquire historical multi-source data, align the historical multi-source data with timestamps, perform time-frequency decomposition, and extract multi-scale feature components. The model building module is used to analyze the causal relationship path between meteorological data, unit operating parameters and power generation based on the multi-scale feature components and the causal discovery algorithm, form a causal relationship network, screen key variables based on the causal relationship network, and build and train a power generation prediction model. The power generation prediction module is used to predict future weather sequences using a meteorological prediction model, perform causal path propagation deduction based on the future weather sequences, generate a deduction sequence of unit operating parameters, and then input the future weather sequences and the deduction sequence of unit operating parameters into the power generation prediction model to obtain the predicted power generation. The verification and correction module is used to verify the physical boundary constraints of the predicted power generation based on the upper and lower limits of the unit power curve corresponding to the current wind speed, correct abnormal prediction values, and obtain the final predicted power generation. The power generation calculation module is used to calculate the predicted power generation for future periods based on the final predicted power generation through time integration.