Wind power multi-time scale generation power joint prediction method and system
By constructing a joint optimization model of cross-scale aggregate constraints and physical constraints, the problems of inconsistency and poor adaptability in wind power prediction are solved, the consistency and accuracy of wind power prediction at multiple time scales are achieved, and the operating efficiency of wind farms and power grids is optimized.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- KUNMING UNIV OF SCI & TECH
- Filing Date
- 2025-12-28
- Publication Date
- 2026-04-21
AI Technical Summary
Existing wind power prediction technologies are inconsistent across different time scales, lack physical constraints, and have poor model adaptability, resulting in inaccurate prediction results that fail to meet the actual operating conditions of wind farms, increasing operation and maintenance costs and grid operation risks.
The annual, monthly, and daily power generation predictions of wind farms are optimized using PSO-OS-ELM, AFSA-OS-ELM, and WOA-LSTM models. By combining cross-scale aggregation constraints and physical feasibility constraints, consistent and accurate multi-timescale power generation prediction results are generated through a joint optimization problem.
It has achieved consistency and accuracy in wind power prediction results across multiple time scales, reduced grid operating costs, optimized wind farm output plans, and improved the economic benefits of wind farms and the security of the power grid.
Smart Images
Figure CN121906407A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method and system for joint prediction of wind power generation at multiple time scales, belonging to the field of new energy power generation prediction technology. Background Technology
[0002] Wind power, as an important clean energy source, is playing an increasingly vital role in the global energy structure. However, the volatility and intermittency of wind power pose challenges to the stable operation of the power grid. Accurate power generation forecasting is crucial for ensuring the safe and stable operation of the power grid, optimizing power dispatch, improving wind power absorption capacity, and facilitating electricity market transactions.
[0003] Existing wind power forecasting technologies are typically researched and applied for a single time scale (ultra-short-term, short-term, medium-term, or long-term). Currently, for daily or shorter-term forecasts, numerical weather prediction (NWP) and machine learning models such as LSTM are commonly used; for monthly forecasts, seasonal characteristics and statistical models are often combined; and for annual forecasts, climate indicators and historical trend analysis are relied upon more heavily. However, these technologies have certain shortcomings. Forecasting models at different time scales usually operate independently, leading to inconsistencies in the aggregation of forecast results over time (the sum of daily forecasts may not equal the monthly forecast, and the sum of monthly forecasts may not equal the annual forecast). This inconsistency creates contradictions and decision-making difficulties for long-term grid planning, medium-term dispatching, and short-term operation, resulting in a lack of consistency in forecast results across different scales. Existing forecasting models often focus only on forecast accuracy, rarely incorporating the actual physical constraints of wind farm operation (wind turbine rated power, maximum output limit, maintenance plan, wake effect, etc.) into the forecast process, leading to forecast results that may exceed the actual output range and limiting their practicality. Existing models lack robustness and adaptability. The operating environment of wind farms, turbine performance, and the accuracy of numerical weather forecasts change dynamically over time. Existing independent models struggle to adapt quickly, requiring frequent manual adjustments and increasing operation and maintenance costs. Current power forecasting methods lack consistent uncertainty quantification methods across time scales, making it difficult to provide unified probabilistic forecast information. This results in a lack of uncertainty information transmission, impacting grid risk assessment and reserve capacity optimization.
[0004] In view of the above problems, there is an urgent need for a wind power generation prediction method that can effectively integrate prediction information from multiple time scales and ensure its cross-scale consistency and physical feasibility. This method not only has important theoretical value but also has extremely high practical value. Summary of the Invention
[0005] This invention aims to solve the technical problems of inconsistent prediction results at different scales, lack of physical constraints, and poor model adaptability in existing wind power prediction, and provides a method and system for joint prediction of wind power generation at multiple time scales.
[0006] The technical solution of this invention is:
[0007] According to a first aspect of the present invention, a method for joint prediction of wind power generation at multiple time scales is provided, comprising the following steps:
[0008] S1. Obtain the annual-scale prediction feature set, the monthly-scale prediction feature set, and the daily-scale prediction feature set;
[0009] S2. Construct the first, second, and third baseline prediction models; use the annual, monthly, and daily prediction feature sets as inputs to the first, second, and third baseline prediction models, respectively, to generate annual power generation baseline prediction results, monthly power generation baseline prediction results, and daily power generation baseline prediction results.
[0010] S3. Construct cross-scale aggregation constraints;
[0011] S4. Define physical feasibility constraints;
[0012] S5. Construct a joint optimization problem, taking the annual, monthly, and daily baseline prediction results as input, with the objective function being to minimize the weighted squared deviation between the reorganized prediction results at each scale and the corresponding baseline prediction results, and taking the cross-scale aggregation constraint and the physical feasibility constraint as hard constraints.
[0013] S6. Solve the joint optimization problem to obtain the final reorganized annual, monthly, and daily power generation joint prediction results that satisfy cross-scale consistency and conform to physical operating logic, and output them.
[0014] Furthermore, the first, second, and third baseline prediction models in S2 are specifically as follows:
[0015] The first baseline prediction model uses the PSO-OS-ELM model, and the particle swarm optimization algorithm is used to optimize the key parameters of the online sequence extreme learning machine.
[0016] The second baseline prediction model uses the AFSA-OS-ELM model, which utilizes the global optimization capability of the artificial fish swarm algorithm to optimize the key parameters of the online sequence extreme learning machine.
[0017] The third baseline prediction model uses the WOA-LSTM model and optimizes the hyperparameters of the long short-term memory neural network using the whale optimization algorithm.
[0018] Furthermore, the cross-scale aggregation constraint is used to describe the mathematical identity relationship between the aggregation of daily forecast results to monthly forecast results and the aggregation of monthly forecast results to annual forecast results.
[0019] Furthermore, the mathematical expression for the cross-scale aggregation constraint is:
[0020] ;
[0021] ;
[0022] In the formula, , These are aggregation matrices from daily to monthly and from monthly to yearly scales, respectively. , , These are the daily, monthly, and annual power generation reforming prediction results to be solved.
[0023] Furthermore, the physical feasibility constraints include real-time dynamic available power limits and non-negative output limits for wind farms.
[0024] Furthermore, the objective function is expressed as:
[0025]
[0026] in, For the solution to be found Scale-based power generation reforming prediction results; , These represent year, month, and day, respectively. for Baseline forecast results for scaled power generation; The preset prediction bias weighting factors for each scale; The smoothness coefficient; This represents the difference in predicted power between adjacent time points on a diurnal scale.
[0027] According to a second aspect of the present invention, a wind power multi-timescale power generation joint prediction system is provided, comprising modules of any of the methods described above.
[0028] According to a third aspect of the present invention, a terminal device is provided, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, the processor being configured to perform the steps of the method described in any one of the preceding inventions.
[0029] The beneficial effects of this invention are as follows: This invention effectively coordinates the relationship between annual, monthly, and daily scale predictions through a cross-scale reorganization and optimization mechanism, ensuring the consistency of prediction results in time aggregation and reducing the accumulation of cross-scale errors; furthermore, it embeds the actual physical constraints of wind farm operation into the reorganization and optimization process, ensuring that the final prediction results conform to actual operating conditions, thereby improving the overall prediction accuracy; through more accurate and consistent predictions, the reserve capacity required for grid operation can be effectively reduced, and the power output plan of wind farms can be optimized, thereby reducing grid operating costs and improving the economic benefits of wind farms. Attached Figure Description
[0030] Figure 1 This is a flowchart of the present invention.
[0031] Figure 2 This is a comparison chart of annual-scale total power generation capacity reforming based on Embodiment 2 of the present invention.
[0032] Figure 3 This is a comparison of monthly forecast results based on the monthly scale provided in Embodiment 2 of the present invention.
[0033] Figure 4 This is based on the details of the diurnal power curve and physical constraint correction provided in Embodiment 2 of the present invention.
[0034] Figure 5 This is based on the multi-scale prediction conflict reorganization and normalization analysis provided in Embodiment 2 of the present invention. Detailed Implementation
[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.
[0036] Example 1: As Figure 1 As shown, according to a first aspect of the present invention, a method for joint prediction of wind power generation at multiple time scales is provided, comprising:
[0037] S1. Obtain the annual-scale prediction feature set, the monthly-scale prediction feature set, and the daily-scale prediction feature set;
[0038] The annual-scale prediction feature set, monthly-scale prediction feature set, and daily-scale prediction feature set are obtained specifically as follows:
[0039] Acquire historical multi-source data; the multi-source data includes wind farm operation data and multi-model numerical weather prediction data (NWP data).
[0040] The multi-source data is preprocessed and multi-scale feature extraction is performed to obtain feature sets for prediction at the annual, monthly, and daily scales, respectively; that is, the multi-source data is cleaned, standardized, missing value imputed, and feature extracted.
[0041] For example, the annual-scale prediction feature set includes annual wind farm operation data and annual weather forecast data. The monthly-scale prediction feature set includes monthly wind farm operation data and monthly weather forecast data. The daily-scale prediction feature set includes daily wind farm operation data and daily weather forecast data.
[0042] S2. Construct the first, second, and third baseline prediction models; use the annual, monthly, and daily prediction feature sets as inputs to the first, second, and third baseline prediction models, respectively, to generate annual power generation baseline prediction results, monthly power generation baseline prediction results, and daily power generation baseline prediction results.
[0043] The first baseline prediction model uses the PSO-OS-ELM model, and the Particle Swarm Optimization (PSO) algorithm is used to optimize the key parameters of the online Sequence Extreme Learning Machine (OS-ELM).
[0044] The second baseline prediction model adopts the AFSA-OS-ELM model, which utilizes the global optimization capability of the Artificial Fish Swarm Algorithm (AFSA) to optimize the key parameters (input weight matrix a and hidden layer bias matrix b) of the online sequence extreme learning machine (OS-ELM).
[0045] The third baseline prediction model uses the WOA-LSTM model, which optimizes the hyperparameters of the Long Short-Term Memory Neural Network (LSTM) using the Whale Optimization Algorithm (WOA).
[0046] As can be seen from the above, based on the monthly feature set, the OS-ELM model is optimized using AFSA (Artificial Fish Swarm Algorithm) to train a monthly baseline prediction model and generate a monthly power generation baseline prediction result; based on the daily feature set, the LSTM (Long Short-Term Memory Neural Network) model is optimized using WOA (Whale Optimization Algorithm) to train a daily baseline prediction model and generate a daily power generation baseline prediction result.
[0047] S3. Construct cross-scale aggregation constraints, which are used to describe the mathematical identity relationship between daily forecast results aggregated to monthly forecast results and monthly forecast results aggregated to annual forecast results.
[0048] S4. Define physical feasibility constraints, including real-time dynamic available power limits and non-negative output limits for wind farms.
[0049] S5. Construct a joint optimization problem, taking the annual, monthly, and daily baseline prediction results as input, with the objective function being to minimize the weighted squared deviation between the reorganized prediction results at each scale and the corresponding baseline prediction results, and taking the cross-scale aggregation constraint and the physical feasibility constraint as hard constraints.
[0050] S6. Solve the joint optimization problem to obtain the final reorganized annual, monthly, and daily power generation joint prediction results that satisfy cross-scale consistency and conform to physical operating logic, and output them.
[0051] According to a second aspect of the present invention, a wind power multi-timescale power generation joint prediction system is provided, comprising modules of any of the methods described above. Each module in the aforementioned wind power multi-timescale power generation joint prediction system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.
[0052] According to a third aspect of the present invention, a terminal device is provided, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, the processor being configured to perform the steps of the method described in any one of the preceding inventions.
[0053] Example 2: Figures 1-5 As shown, a method for joint prediction of wind power generation across multiple time scales includes the following:
[0054] I. Multi-source data acquisition and preprocessing steps.
[0055] By integrating data interfaces from wind farm SCADA systems and multiple numerical weather prediction (NWP) service providers, real-time or periodic acquisition of multi-source data for the target area is achieved. The collected data types cover wind turbine operating parameters (power, wind speed, wind direction, temperature) and weather forecast data (wind speed, temperature, air pressure, humidity, etc.). After data acquisition, data cleaning, standardization, missing value imputation, and multi-scale feature extraction are performed to ensure the accuracy and consistency of the input data.
[0056] II. Steps for constructing and training a multi-scale baseline prediction model.
[0057] Three independent baseline prediction models were constructed to handle prediction tasks at different time scales, and these models were enhanced by optimization algorithms.
[0058] For annual-scale power generation forecasting, an annual baseline forecasting model (i.e., the first baseline forecasting model) is constructed and trained. This model employs the PSO-OS-ELM model, where the Particle Swarm Optimization (PSO) algorithm is used to optimize the key parameters of the Online Sequential Extreme Learning Machine (OS-ELM). The core idea of OS-ELM is to randomly generate the weights and biases between the input layer and the hidden layer, and to solve for the weights between the hidden layer and the output layer using the Moore-Penrose generalized inverse. For a given input... The OS-ELM of the activation function g(x), its output It can be represented as:
[0059] ;
[0060] In the formula, This represents the number of neurons in the hidden layer. For the first The input weight vector of each hidden layer neuron. To bias it, For the first The weights from each hidden layer neuron to the output layer. The PSO algorithm uses the input weight matrix of OS-ELM. and hidden layer bias matrix The constructed vector serves as the position of the individual particle. The root mean square error (RMSE) of the OS-ELM prediction model is used as the fitness function for optimization. The formula for calculating the root mean square error is as follows:
[0061] ;
[0062] In the formula, n is the number of training samples for the OS-ELM prediction model. This indicates the predicted wind power generation capacity, and This represents the actual wind power generation. PSO iteratively searches for the optimal (a,b) combination that minimizes the RMSE. During the iteration process, each particle... Based on its individual historical best position and the group's historical best position Update its speed and location The particle velocity update formula is:
[0063] ;
[0064] The particle position update formula is:
[0065] ;
[0066] In the formula, For the number of iterations, For inertial weights, , As a learning factor, , The value is a random number between [0,1]. The OS-ELM model determines the input weights, biases, and initial output weights during the initialization phase. Its online learning capability enables it to quickly update the output weights based on newly added training sample data and continuously optimize the overall model parameters with the assistance of PSO, thereby generating a baseline prediction of annual power generation.
[0067] For monthly power generation forecasting, a monthly baseline forecasting model (i.e., the second baseline forecasting model) is constructed and trained. This model employs the AFSA-OS-ELM model, utilizing the global optimization capability of the Artificial Fish Swarm Algorithm (AFSA) to optimize the input weight matrix 'a' and the hidden layer bias matrix 'b' of the OS-ELM. The AFSA algorithm explores and searches the parameter space to find the prediction cost function of OS-ELM by simulating the foraging, swarming, and tail-chasing behaviors of artificial fish. The minimum optimal solution. The objective function of the AFSA-OS-ELM can be expressed as:
[0068] ;
[0069] in, Let a and b be the input weights and bias, respectively, β be the output weight, H be the hidden layer output matrix, T be the target output matrix, and N be the number of samples. AFSA iteratively calculates the root mean square error of artificial fish individuals at different locations as the food density, updates the optimal state and optimal value of the artificial fish, and finally obtains the optimal input weights and hidden layer bias. The monthly baseline prediction model utilizes a monthly-scale feature set during training. This model also has online learning capabilities, enabling parameter updates and model adaptation based on new monthly data.
[0070] For daily-scale power generation forecasting, a daily baseline prediction model (i.e., the third baseline prediction model) is constructed and trained. This model employs a WOA-LSTM model, where the Whale Optimization (WOA) algorithm is used to optimize the hyperparameters of the Long Short-Term Memory (LSTM) neural network. The WOA algorithm dynamically adjusts the LSTM hyperparameters by simulating whale encirclement and spiral bubble web attack behaviors. During the encirclement phase, the whale's position... The update formula is:
[0071] ;
[0072] ;
[0073] In the formula, This is the current location of the whale. This represents the current optimal position for the whale (prey). and To control parameters. During the exploration phase, whale location updates can be achieved by randomly selecting the locations of other whales. accomplish:
[0074] ;
[0075] The LSTM model, through its gating mechanism, effectively addresses the shortcomings of traditional RNNs in learning long sequences, enabling it to flexibly capture complex patterns in time series. Given an input sequence... and the hidden state of the previous moment The calculation process for each time step of LSTM is as follows:
[0076] Forgotten Gate Decision based on past cell states What information is discarded:
[0077] f t = σ ( W f ⋅ [ h t − 1 , x t ] + b f ) ;
[0078] Input gate Determine the current input and the hidden state of the previous moment How much information is used to update the current cell state while generating candidate cell states? :
[0079] i t = σ ( W i ⋅ [ h t − 1 , x t ] + b i ) ;
[0080] C ̃ t = t a n h ( W C ⋅ [ h t − 1 , x t ] + b C ) ;
[0081] Cell state update By combining the results of the forgetting gate and the input gate, the cell's memory state is updated.
[0082] ;
[0083] Output gate Determine the current cell state How much information is output to the hidden state? :
[0084] o t = σ ( W o ⋅ [ h t − 1 , x t ] + b o ) ;
[0085] Final hidden state for:
[0086] ;
[0087] In the formula, , , , These are the weight matrices for the forget gate, input gate, candidate cell states, and output gate, respectively. For bias vectors, Represents the sigmoid activation function. Represents the hyperbolic tangent activation function. This is an element-wise multiplication. WOA iteratively optimizes to minimize the prediction error (RMSE, MAE) of the LSTM model on the test set.
[0088] III. Steps for constructing cross-scale aggregation constraints and defining physical constraints.
[0089] Cross-scale aggregation constraints are implemented through an aggregation matrix, which defines the strict mathematical relationship between prediction results at different time scales. The sum of the predicted daily power for a given month must equal the monthly predicted power for that month, and the sum of the predicted monthly power for a given year must equal the annual predicted power for that year. These constraints guarantee the self-consistency of the prediction results across the time dimension. Cross-scale aggregation constraints are expressed as:
[0090] ;
[0091] ;
[0092] In the formula, , These are aggregation matrices from daily to monthly and from monthly to yearly scales, respectively. , , These are the daily, monthly, and annual power generation reforming prediction results for the subsequent joint optimization problem to be solved.
[0093] Physical feasibility constraints ensure that the predicted results conform to the actual physical limitations of the wind farm's operation. The predicted power value at any given time cannot exceed the wind farm's real-time dynamic available power and is subject to non-negative output limitations. These physical feasibility constraints are expressed as follows:
[0094] ;
[0095] ;
[0096] In the formula, For real-time dynamic available power, The rated installed capacity of the wind farm For at any time The capacity for downtime maintenance.
[0097] IV. Cross-scale restructuring and optimization steps.
[0098] Using the annual, monthly, and daily baseline prediction results as input, the objective function is to minimize the weighted squared deviation between the renormalized prediction results at each scale and the corresponding baseline prediction results. The cross-scale aggregation constraint and the physical feasibility constraint are used as hard constraints to construct a joint optimization problem. To ensure the smoothness of the prediction results, a penalty term is introduced into the objective function to avoid unnatural and drastic changes in the renormalized predictions. The objective function of the joint optimization problem can be expressed as:
[0099]
[0100] The constraints can be expressed as:
[0101]
[0102] in, The results of power generation reforming prediction at various scales to be solved ( ,when If we choose D, then for (The same applies to other cases); Baseline prediction results for power generation at various scales; The preset prediction bias weighting factors for each scale; The smoothness coefficient; The difference between the predicted power at adjacent times on a daily scale is used to constrain the predicted values at different times to ensure a smooth transition.
[0103] V. Prediction Results Output and Application Steps.
[0104] Finally, this invention invokes a quadratic programming algorithm to find the optimal combination that minimizes the deviation from the original model among all possible outcomes that satisfy the aggregation logic and physical upper bound (capacity limit). The system outputs data through standardized interfaces to the power grid dispatching system, the electricity market trading platform, the power grid's long-term planning department, and the wind farm's own operation and management system. These forecasts can be used for power grid capacity planning, reserve capacity decisions, electricity market bidding strategy formulation, and optimal wind farm operation and dispatch, thereby significantly improving the economic benefits of wind power grid connection and system operational safety. Furthermore, the system can provide additional information such as historical forecasts versus actual values, forecast error analysis, and confidence intervals for forecast results to support more advanced risk management and decision-making.
[0105] The following is a further explanation using simulation data:
[0106] like Figure 2As shown in the figure, the original baseline power generation forecast and the forecast after recalibration according to the present invention are compared on an annual timescale. In this embodiment, the annual baseline (as shown in the left bar chart, the value is 22419.6 MWh) serves as the initial reference for the macro-level power generation target. After recalibration optimization according to the present invention, the generated recalibrated forecast (as shown in the right bar chart, the value is 22509.9 MWh) is finely adjusted by incorporating confidence features at the monthly and daily scales, while acknowledging the initial forecast bias. This figure demonstrates that the present invention can establish a robust annual expected target and provides a macro-level "anchor" for subsequent monthly and daily numerical alignment.
[0107] like Figure 3 As shown in the figure, paired bar charts illustrate the predicted monthly electricity consumption before and after renormalization at a monthly scale. The figure reveals that the original monthly baseline predictions generally exhibit a systematic positive bias (i.e., the predicted values are slightly overestimated). This invention, through cross-scale aggregation equation constraints, mandates that the sum of the 12-month predictions equal the annual renormalization value, driving the model to perform a global correction to the monthly prediction sequence (monthly-scale renormalization prediction). This figure visually demonstrates how this invention eliminates the numerical conflict between monthly and annual data, achieving logical self-consistency in the monthly electricity consumption distribution.
[0108] like Figure 4 As shown in the figure, the daily-scale predicted power variation curve is recorded from day 100 to day 130. The daily-scale baseline prediction marks the original daily-scale baseline prediction without considering the actual operating conditions. During the unit maintenance period from day 105 to day 115, the unit maintenance upper limit represents the physical output upper limit limit (15 MW) for that period. It can be seen that the original baseline prediction significantly exceeds the physical limit, while the reforming curve (daily-scale reforming prediction) output by this invention is precisely constrained within the maintenance upper limit, eliminating physically infeasible solutions. The daily-scale reforming prediction maintains a smooth physical ramp-up trend even when forcibly correcting data mutations caused by maintenance. This proves that this invention, through the sequence smoothing penalty term in the objective function, ensures that the prediction results conform to the dynamic characteristics of the actual operation of the wind turbine.
[0109] like Figure 5As shown in the figure, this further illustrates the core logic of the present invention in resolving "model conflicts." Before rebalancing, due to the independence of models at each scale, the daily total (21946.0 MWh), monthly total (25225.2 MWh), and annual baseline predictions exhibited significant algebraic inconsistencies, indicating a serious data "conflict." After applying the method of this invention (after rebalancing), the algorithm performed the following corrections: the daily total was revised upwards by 2.4% to compensate for the conservative bias of the daily scale model; the monthly total was revised downwards by a substantial 10.9% to offset the systematic overestimation of the monthly model; and the annual scale was fine-tuned and anchored by 0.4%. After rebalancing, the predicted total electricity volume in the three dimensions aligned with the "final consistency target" line. This figure powerfully demonstrates that the present invention can utilize cross-scale joint optimization to force global normalization of multi-time-scale prediction results, providing data support with logically closed-loop characteristics for power system scheduling.
[0110] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.
Claims
1. A method for joint prediction of wind power generation capacity across multiple time scales, characterized in that, Includes the following steps: S1. Obtain the annual-scale prediction feature set, the monthly-scale prediction feature set, and the daily-scale prediction feature set; S2. Construct the first, second, and third baseline prediction models; use the annual, monthly, and daily prediction feature sets as inputs to the first, second, and third baseline prediction models, respectively, to generate annual power generation baseline prediction results, monthly power generation baseline prediction results, and daily power generation baseline prediction results. S3. Construct cross-scale aggregation constraints; S4. Define physical feasibility constraints; S5. Construct a joint optimization problem, taking the annual, monthly, and daily baseline prediction results as input, with the objective function being to minimize the weighted squared deviation between the reorganized prediction results at each scale and the corresponding baseline prediction results, and taking the cross-scale aggregation constraint and the physical feasibility constraint as hard constraints. S6. Solve the joint optimization problem to obtain the final reorganized annual, monthly, and daily power generation joint prediction results that satisfy cross-scale consistency and conform to physical operating logic, and output them.
2. The wind power multi-timescale joint prediction method according to claim 1, characterized in that, The first, second, and third baseline prediction models in S2 are specifically as follows: The first baseline prediction model uses the PSO-OS-ELM model, and the particle swarm optimization algorithm is used to optimize the key parameters of the online sequence extreme learning machine. The second baseline prediction model uses the AFSA-OS-ELM model, which utilizes the global optimization capability of the artificial fish swarm algorithm to optimize the key parameters of the online sequence extreme learning machine. The third baseline prediction model uses the WOA-LSTM model and optimizes the hyperparameters of the long short-term memory neural network using the whale optimization algorithm.
3. The wind power multi-timescale joint prediction method according to claim 1, characterized in that, The cross-scale aggregation constraint is used to describe the mathematical identity relationship between the aggregation of daily forecast results to monthly forecast results and the aggregation of monthly forecast results to annual forecast results.
4. The wind power multi-timescale joint prediction method according to claim 3, characterized in that, The mathematical expression for the cross-scale aggregation constraint is: ； ； In the formula, , These are aggregation matrices from daily to monthly scale and from monthly to yearly scale, respectively; , , These are the daily, monthly, and annual power generation reforming prediction results to be solved.
5. The method for joint prediction of wind power generation capacity across multiple time scales according to claim 1, characterized in that, The physical feasibility constraints include the real-time dynamic available power limit of the wind farm and the non-negative output limit.
6. The wind power multi-timescale joint prediction method according to claim 1, characterized in that, The objective function is expressed as follows: ； in, For the solution to be found Scale-based power generation reforming prediction results; , These represent year, month, and day, respectively. for Baseline forecast results for scaled power generation; The preset prediction bias weighting factors for each scale; The smoothness coefficient; This represents the difference in predicted power between adjacent time points on a diurnal scale.
7. A wind power multi-timescale power generation joint prediction system, characterized in that, The module includes the wind power multi-timescale power generation joint prediction method according to any one of claims 1-6.
8. A terminal device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, The processor is configured to perform the steps of the method according to any one of claims 1-6.