Weather intelligence-based fan output prediction method and system

Abstract

The invention discloses a wind turbine output prediction method and system based on meteorological intelligence, and relates to the technical field of power generation. Preprocessing the collected data, and automatically determining a resampling standard according to a time sequence mode and a change trend of the data; segmenting the data set by comprehensively considering the time sequence, geographic and climate characteristics of the data; training the model by using an LSTM and a CNN fused with the LSTM so as to improve the recognition capability of a long-term meteorological mode and the capture capability of spatial distribution characteristics of a wind power plant; inputting real-time or latest collected weather forecast data and wind power plant operation condition data by using the trained model, and executing fan output prediction; and evaluating the operation effect of the wind power plant according to the prediction result to guide equipment maintenance and transformation. According to the invention, by constructing an end-to-end meteorological intelligent system, deep understanding and high-quality feature extraction of complex meteorological data are realized.

Description

Technical Field

[0001] This invention relates to the field of power generation technology, and in particular to a method and system for predicting wind turbine output based on meteorological intelligence. Background Technology

[0002] In the rapid development of the wind energy industry, wind turbine output forecasting technology plays a crucial role. Traditional wind turbine output forecasting methods are mainly based on historical data and statistical models, combined with basic meteorological elements such as wind speed and direction. This method has been widely used in wind turbine output forecasting, but it still has certain limitations.

[0003] First, traditional forecasting techniques are inefficient at processing large-scale, high-dimensional meteorological data and struggle to accurately capture the dynamics of complex climate systems. Due to the instability of wind power generation and its high dependence on meteorological conditions, traditional methods often fail to fully meet the accuracy requirements of practical applications. For example, traditional models struggle to accurately predict sudden changes in wind speed and wind turbine output under extreme weather conditions.

[0004] Secondly, due to the lack of effective data processing and analysis tools, traditional forecasting methods are significantly inefficient in data utilization. Existing technologies often cannot fully utilize the massive meteorological big data resources, resulting in inaccurate forecasts and an inability to provide effective support for the operation and maintenance management of wind farms.

[0005] Furthermore, the limitations of traditional technologies in handling nonlinear and non-stationary meteorological processes restrict their application under complex climatic conditions. Predicting wind turbine output requires considering the interaction of multiple factors, such as topography and climate change, but existing technologies often fail to fully account for these factors, leading to increased uncertainty in forecast results. Summary of the Invention

[0006] In view of the problems of low accuracy, inflexible model updates, and inability to adapt to environmental changes in existing wind turbine output prediction technologies and systems, this invention is proposed.

[0007] Therefore, the problem to be solved by this invention is how to construct an end-to-end, ubiquitous intelligent power output prediction system for wind farms, so as to achieve a deep understanding and dynamic modeling of complex meteorological environments, thereby continuously improving wind power utilization efficiency.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] In a first aspect, embodiments of the present invention provide a wind turbine output prediction method based on meteorological intelligence, which includes collecting multi-source data including historical wind turbine output data, climate characteristic data, meteorological forecast data, and topographic data; preprocessing the collected data and automatically determining the resampling standard based on the time series pattern and trend of the data; segmenting the dataset by comprehensively considering the time sequence, geographical and climatic characteristics of the data; training the model using LSTM and CNN fused with LSTM respectively to improve the ability to recognize long-term meteorological patterns and capture the spatial distribution characteristics of wind farms; using the trained model, inputting real-time or latest collected meteorological forecast data and wind farm operation status data to perform wind turbine output prediction; and evaluating the wind farm operation effect based on the prediction results to guide equipment maintenance and renovation.

[0010] As a preferred embodiment of the wind turbine output prediction method based on meteorological intelligence described in this invention, the automatic determination of resampling criteria based on the time series pattern and trend of the data includes the following steps: constructing a flexible and scalable data import module to support real-time acquisition of multi-source heterogeneous time series data; using an unsupervised anomaly detection algorithm based on isolated forests to automatically identify low-quality data, and using data augmentation and generative models to automatically complete missing values; using an unsupervised binary clustering algorithm based on tree models to evaluate data distribution in real time and select the optimal normalization scheme; learning high-order time-dependent features of the data through an automatic feature learning module based on deep neural networks; using an autoregressive model based on Fourier and wavelet transforms to evaluate data periodicity in real time and dynamically adjust the resampling interval; employing a Bayesian optimization search algorithm to search online for the optimal preprocessing parameters and module combinations; utilizing feedback from downstream prediction tasks to fine-tune each preprocessing pipeline through multi-task learning end-to-end; using a locally interpretable model to generate explanations for the decisions of each preprocessing step in real time; and employing online learning and incremental learning algorithms to enable the preprocessing model to continuously adapt to data changes.

[0011] As a preferred embodiment of the wind turbine output prediction method based on meteorological intelligence described in this invention, the dataset segmentation includes the following steps: applying graph convolutional networks to model potential relationships between regions to enrich the contextual representation of time-series data and capture the spatiotemporal correlation between data; constructing a knowledge graph and fusing multi-source heterogeneous geographical environment data to perform external knowledge transfer and inject additional semantic expression into the original dataset; constructing multi-time-granularity ensemble sub-models, selecting a coarser-grained model if the time series belongs to a low-frequency periodic pattern, and selecting a finer-grained model if it reflects high-frequency data change characteristics, to comprehensively simulate various potential periodicities of the time series; encapsulating a multi-agent adversarial search mechanism to formalize the dataset segmentation problem into a decision-making process, and finding robust segmentation strategies through game theory; identifying similar and comparable clusters in the dataset, and applying transfer learning algorithms to perform adaptive knowledge transfer between clusters to improve the segmentation effect; and constructing an agent environment based on reinforcement learning theory and continuously generating ensemble feedback to make the segmentation model evolve towards a better solution.

[0012] As a preferred embodiment of the meteorological intelligence-based wind turbine output prediction method described in this invention, the training of the model using LSTM and a fused LSTM CNN includes the following steps: integrating multi-source data through a data fusion framework and creating a comprehensive dataset through feature selection algorithms and generative feature construction; customizing and improving the LSTM model architecture to adapt to the temporal complexity of meteorological data; developing an improved CNN architecture and using multi-scale convolutional kernels to conduct in-depth analysis of the spatial distribution characteristics of wind farms; constructing a fusion model through weighted fusion and model stacking to integrate the prediction capabilities of LSTM and CNN; determining whether the fusion model needs fine-tuning based on real-time data and historical performance data; establishing a comprehensive performance evaluation system to comprehensively evaluate model performance through prediction interval width or coverage probability; and establishing a continuous model iteration and update mechanism based on the performance evaluation results to dynamically optimize the model structure and parameters to adapt to environmental changes.

[0013] As a preferred embodiment of the wind turbine output prediction method based on meteorological intelligence described in this invention, the customized improvement of the LSTM model architecture includes the following steps: constructing a multi-layer LSTM structure by increasing the number of layers in the LSTM network to enhance the model's ability to capture and learn long-term time-series data dependencies; implementing a dynamic neuron number configuration mechanism in each LSTM layer and determining the optimal number of neurons in each layer through experimental analysis; embedding an attention mechanism in the LSTM structure to improve the LSTM network's ability to identify and weight key time-series features; adopting a bidirectional LSTM structure to simultaneously process the forward and reverse information flows of the sequence and integrate past and future contextual information; applying Dropout operations and L1 / L2 regularization strategies in the LSTM network, and preventing overfitting of the LSTM model by randomly discarding some network connections and applying regularization penalties to the weights; implementing a dynamic learning rate adjustment mechanism, automatically adjusting the learning rate using an adaptive learning rate optimization algorithm; the specific formula for the customized improved LSTM unit is as follows:

[0014] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0015] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0016]

[0017]

[0018] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0019] h t =o t *tanh(c t )

[0020] Where σ represents the Sigmoid function, f t i t and o t These represent the forget gate, input gate, and output gate, respectively. t Represents the cell state, and tanh represents the hyperbolic tangent activation function. The candidate cell state is represented by *, element-wise multiplication is represented by W and b, which represent the weight matrix and bias term, respectively. The subscripts f, i, c, and o correspond to the forget gate, input gate, cell state, and output gate, respectively. t h represents the input at a given time t. t-1 This represents the hidden state at the previous time step t-1.

[0021] As a preferred embodiment of the wind turbine output prediction method based on meteorological intelligence described in this invention, the development and improvement of the CNN architecture includes the following steps: using convolutional kernels of different sizes in parallel at the same level of the CNN model, enabling the model to simultaneously capture local details and global features in the wind farm; automatically identifying and concentrating CNN model resources by applying a spatial attention weighting mechanism to the convolutional feature maps, focusing on key spatial features for wind power output prediction; applying depthwise separable convolution technology to reduce the complexity of the CNN model network and the consumption of computational resources; developing a dynamic convolutional kernel learning strategy, allowing the CNN model to adaptively adjust the size and shape of the convolutional kernels according to the specific wind farm layout and environmental conditions; and implementing feature pyramid networks or cross-layer connection technology to fuse feature information from different levels.

[0022] As a preferred embodiment of the wind turbine output prediction method based on meteorological intelligence described in this invention, the following steps are included: adaptively adjusting the size and shape of the convolution kernel according to the specific wind farm layout and environmental conditions: constructing a real-time monitoring system by monitoring the data flow of the wind farm to determine the specific operating conditions of the wind farm; constructing decision rules: if the wind speed is consistently higher than the average wind speed and fluctuates little, a smaller convolution kernel is selected to focus on capturing high-resolution local features to improve the accuracy of the prediction details; if the wind speed changes drastically in a short period of time, the convolution kernel is dynamically adjusted to make its shape more extensive and span multiple data points to capture sudden spatial variations; if the wind direction data shows regular periodic changes, an L-shaped or J-shaped convolution kernel is used to lock the pattern of periodic wind direction changes; if multiple turbulent superposition areas are detected in the wind farm, a cross shape or a star shape is applied. The system employs various convolutional kernels, focusing on multiple intersection points to analyze the superposition of multiple patterns. If the wind turbine layout within the wind farm is non-uniform, resulting in highly uneven local wind flow distribution, irregularly shaped or refined convolutional kernels are used to meticulously capture local wind flow characteristics. Real-time monitoring of various indicators of the wind farm is captured, and preliminary analysis is performed through a real-time monitoring system to assess the current environmental state. Based on the data analyzed by the real-time monitoring system, decision rules are used to dynamically adjust the convolutional kernels to capture key features. The adjusted CNN model predicts wind power output and compares it with actual measurements to calculate the prediction error. The source of the prediction error is analyzed to determine if it is due to improper convolutional kernel adjustment. Based on the error analysis results, the parameters of the convolutional kernel adjustment rules and decision rules are adjusted and optimized. When the system detects that the increase in prediction accuracy is slowing down or reaches a predetermined standard, the optimization loop is completed.

[0023] Secondly, embodiments of the present invention provide a wind turbine output prediction system based on meteorological intelligence, which includes a multi-source data collection module for collecting multi-source data including historical wind turbine output data, climate characteristic data, meteorological forecast data, and topographic data; a preprocessing module for preprocessing the collected data and automatically determining resampling criteria based on the time series pattern and trend of the data; a segmentation module for segmenting the dataset by comprehensively considering the time sequence, geographical and climatic characteristics of the data; a training module for training the model using LSTM and an improved CNN respectively to improve the ability to recognize long-term meteorological patterns and capture the spatial distribution characteristics of wind farms; a prediction module for using the trained model, inputting real-time or latest collected meteorological forecast data and wind farm operation status data, to predict wind turbine output; and an evaluation module for evaluating the wind farm operation effect based on the prediction results to guide equipment maintenance and renovation.

[0024] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the wind turbine output prediction method based on meteorological intelligence as described in the first aspect of the present invention are implemented.

[0025] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, they implement the steps of the wind turbine output prediction method based on meteorological intelligence as described in the first aspect of the present invention.

[0026] The beneficial effects of this invention are as follows: By constructing an end-to-end intelligent meteorological system, this invention achieves deep understanding and high-quality feature extraction of complex meteorological data; by designing a model with autonomous learning and adaptive capabilities, it enables real-time and continuous prediction of wind farms without human intervention; by evaluating operational effectiveness based on prediction results, it guides equipment maintenance and upgrades, reduces costs, and improves economic efficiency; by integrating various intelligent algorithms into the end-to-end system, it achieves perception, analysis, decision-making, and control of the entire wind farm; and by improving the economic efficiency and reliability of wind power, it accelerates the pace of wind power replacing traditional energy sources and leverages the environmental and social benefits of wind power. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating the steps of a meteorological intelligence-based wind turbine output prediction method.

[0029] Figure 2 This is a diagram of a computer device for predicting wind turbine output based on meteorological intelligence. Detailed Implementation

[0030] 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.

[0031] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0032] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0033] Example 1

[0034] Reference Figures 1-2 This is the first embodiment of the present invention, which provides a wind turbine output prediction method based on meteorological intelligence, including:

[0035] S1: Collect multi-source data, including historical wind turbine output data, climate characteristic data, weather forecast data, and topographic data.

[0036] Specifically, multi-source data is collected, including historical wind turbine output data, climate characteristic data, meteorological forecast data, and topographic data. This data not only covers the historical operation of wind farms, but also includes predictions of future meteorological conditions and key information on the impact of topography on wind speed, providing a comprehensive data foundation for model training and prediction.

[0037] S2: Preprocess the collected data and automatically determine the resampling standard based on the time series pattern and trend of the data.

[0038] Specifically, a flexible and scalable data import module is constructed to support real-time acquisition of multi-source heterogeneous time-series data; an unsupervised anomaly detection algorithm based on isolated forests is used to automatically identify low-quality data, and data augmentation and generative models are used to automatically fill in missing values; an unsupervised binary clustering algorithm based on tree models is used to evaluate data distribution in real time and select the optimal normalization scheme; an automatic feature learning module based on deep neural networks is used to learn the high-order time-dependent features of the data; an autoregressive model based on Fourier and wavelet transforms is used to evaluate data periodicity in real time and dynamically adjust the resampling interval; a Bayesian optimization search algorithm is used to search online for the optimal preprocessing parameters and module combinations; feedback from downstream prediction tasks is used to fine-tune the preprocessing pipeline end-to-end through multi-task learning; a locally interpretable model is used to generate explanations for the decisions of each preprocessing step in real time; and online learning and incremental learning algorithms are used to enable the preprocessing model to continuously adapt to data changes.

[0039] S3: The dataset is segmented by comprehensively considering the temporal order, geographical and climatic characteristics of the data.

[0040] Preferably, convolutional networks are applied to model potential relationships between regions to enrich the contextual representation of time-series data and capture the spatiotemporal correlations between data; a knowledge graph is constructed to integrate multi-source heterogeneous geographical environment data for external knowledge transfer, injecting additional semantic expression into the original dataset; multi-temporal granularity ensemble sub-models are constructed, with a coarser-grained model selected if the time series belongs to a low-frequency periodic pattern, and a finer-grained model selected if it reflects high-frequency data change characteristics, to comprehensively simulate various potential periodicities of the time series; similar and comparable clusters in the dataset are identified, and transfer learning algorithms are applied to adaptively transfer knowledge between clusters to improve the segmentation effect; based on reinforcement learning theory, a proxy environment is constructed and ensemble feedback is continuously generated, enabling the segmentation model to evolve towards a better solution.

[0041] S4: The model was trained using LSTM and CNN fused with LSTM respectively to improve the ability to identify long-term meteorological patterns and capture the spatial distribution characteristics of wind farms.

[0042] Specifically, it includes the following steps:

[0043] S4.1: Integrate multi-source data through a data fusion framework, and create a comprehensive dataset through feature selection algorithms and generative feature construction.

[0044] Preferably, an advanced data fusion framework is implemented to integrate multi-source data from satellite remote sensing, ground monitoring stations, and meteorological stations. Automated feature engineering techniques, including but not limited to feature selection algorithms and generative feature construction, are employed to extract and build the most effective feature set required for predictive models from the comprehensive dataset.

[0045] S4.2: Customize and improve the LSTM model architecture to adapt to the temporal complexity of meteorological data.

[0046] Specifically, a multi-layer LSTM structure is constructed by increasing the number of layers in the LSTM network to enhance the model's ability to capture and learn dependencies in long-term temporal data. A dynamic neuron number configuration mechanism is implemented in each LSTM layer, and the optimal number of neurons in each layer is determined through experimental analysis. An attention mechanism is embedded in the LSTM structure to improve the LSTM network's ability to identify and weight key temporal features. A bidirectional LSTM structure is adopted to simultaneously process the forward and reverse information flows of the sequence and integrate past and future contextual information. Dropout operations and L1 / L2 regularization strategies are applied in the LSTM network, and overfitting of the LSTM model is prevented by randomly discarding some network connections and applying regularization penalties to the weights. A dynamic learning rate adjustment mechanism is implemented, automatically adjusting the learning rate using an adaptive learning rate optimization algorithm.

[0047] It should be noted that the customized and improved LSTM cell formula is as follows:

[0048] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0049] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0050]

[0051]

[0052] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0053] h t =o t *tanh(c t )

[0054] Where σ represents the Sigmoid function, f t i t and o t These represent the forget gate, input gate, and output gate, respectively. t Represents the cell state, and tanh represents the hyperbolic tangent activation function. The candidate cell state is represented by *, element-wise multiplication is represented by W and b, which represent the weight matrix and bias term, respectively. The subscripts f, i, c, and o correspond to the forget gate, input gate, cell state, and output gate, respectively. t h represents the input at a given time t. t-1 This represents the hidden state at the previous time step t-1.

[0055] It should be noted that the weighted formula for the attention mechanism is as follows:

[0056] a t =Attention(h t-1 ,x t )

[0057] h t =o t *tanh(c t *a t )

[0058] Among them, c t Represents the cell state, tanh represents the hyperbolic tangent activation function, and a t Represents attention weights, o t This represents the output gate, indicating the output based on the current input x. t and the previous hidden state h t-1 The attention weights, h, are calculated. t Indicates the current hidden state.

[0059] It should be noted that the specific formula for bidirectional LSTM is as follows:

[0060]

[0061] Among them, h t Indicates the current hidden state. Indicates positive output. This indicates reverse output.

[0062] S4.3: Develop an improved CNN architecture and use multi-scale convolution kernels to conduct in-depth analysis of the spatial distribution characteristics of wind farms.

[0063] Preferably, convolutional kernels of different sizes are used in parallel at the same level of the CNN model, enabling the model to capture both local details and global features in the wind farm simultaneously; by applying a spatial attention weighting mechanism to the convolutional feature maps, the CNN model resources are automatically identified and concentrated, focusing on the key spatial features for wind power output prediction; depthwise separable convolution technology is applied to reduce the complexity of the CNN model network and the consumption of computational resources; a dynamic convolutional kernel learning strategy is developed, allowing the CNN model to adaptively adjust the size and shape of the convolutional kernels according to the specific wind farm layout and environmental conditions; and feature pyramid networks or cross-layer connection technology are implemented to fuse feature information from different levels.

[0064] Specifically, adaptively adjusting the size and shape of the convolution kernel based on the specific wind farm layout and environmental conditions includes the following steps: A real-time monitoring system is built by monitoring the wind farm's data stream to determine the specific operating conditions of the wind farm; if the wind speed is consistently higher than the average wind speed with small fluctuations, a smaller convolution kernel is selected to focus on capturing high-resolution local features and improve the accuracy of prediction details; if the wind speed changes drastically in a short period, the convolution kernel is dynamically adjusted to expand its shape (e.g., from a square to a rectangle) and span multiple data points to capture sudden spatial variations; if the wind direction data shows regular periodic changes (possibly due to airflow patterns formed by terrain or buildings), an L-shaped or J-shaped convolution kernel is used to lock in the pattern of periodic wind direction changes; if the relationships between various turbulences in the wind farm become complex (e.g., multiple turbulence superposition areas appear), a cross shape is applied. The system employs either star-shaped convolutional kernels, focusing on multiple intersection points to analyze the superposition of multiple patterns; if the wind turbine layout within the wind farm is non-uniform, resulting in highly uneven local wind flow distribution, irregular or refined convolutional kernels (such as cross-shaped or ring-shaped kernels) are used to meticulously capture local wind flow characteristics; various monitoring indicators of the wind farm are captured in real time, and preliminary analysis is performed through a real-time monitoring system to assess the current environmental state; based on the data analyzed by the real-time monitoring system, decision rules are used to dynamically adjust the convolutional kernels to capture key features; the adjusted CNN model predicts wind power output and compares it with actual measurements to calculate the prediction error; the source of the prediction error is analyzed to determine whether it is due to improper adjustment of the convolutional kernel; based on the error analysis results, the parameters of the convolutional kernel adjustment rules and decision rules are adjusted and optimized; when the system detects that the increase in prediction accuracy slows down or reaches a predetermined standard, the optimization cycle is completed.

[0065] It should be noted that the formula used for CNN structure adjustment is as follows:

[0066] Conv terrain =ReLU(Conv(W) terrain *X)+b terrain )

[0067] Among them, W terrain This represents the convolution kernel weights specifically designed for terrain features, b terrain This represents the corresponding bias term, and X represents the input wind farm data.

[0068] S4.4: Construct a fusion model by weighted fusion and model stacking to integrate the predictive capabilities of LSTM and CNN.

[0069] Preferably, a weighted fusion mechanism is designed to optimize the weighted merging of the prediction outputs of LSTM and CNN models; a stacking strategy is used to take the prediction results of LSTM and CNN models as input to the secondary model; a dynamic fusion framework is developed to dynamically adjust the parameters of LSTM and CNN models in weighted fusion and model stacking based on real-time performance feedback to adapt to changing climate conditions and different wind farm configurations over time; a particle swarm optimization algorithm is integrated to find the optimal weighted fusion parameters and stacked model parameter configurations by simulating natural selection or swarm intelligence behavior; the prediction performance of the fusion model is comprehensively evaluated, and the fusion strategy is finely adjusted based on the evaluation results to ensure that the fusion model maintains optimal prediction performance in different scenarios.

[0070] It should be noted that the secondary model (or meta-model) is responsible for learning the combination of different model outputs to produce the final prediction.

[0071] S4.5: Determine whether the fusion model needs fine-tuning based on real-time data and historical performance data.

[0072] Preferably, a highly adaptive data monitoring system is developed to collect and analyze meteorological data and wind turbine performance data from wind farms in real time; a historical data backtracking analysis module is designed to periodically evaluate the historical prediction performance of the fusion model using machine learning algorithms to identify and learn the changing trends of model performance and the adjustment space of weights; a multi-factor decision-making mechanism and machine learning algorithms are applied to determine whether fine-tuning of the fusion model is necessary; if fine-tuning is determined, parameter optimization, structural updates, or retraining of the model are triggered, and the principle of minimal intervention is implemented to reduce unnecessary resource consumption; the fine-tuned fusion model is validated using an independent test set or validation set, and the performance indicators before and after fine-tuning are compared; based on the evaluation of the fine-tuning effect, the fine-tuning strategy and rules are continuously iterated and optimized.

[0073] It should be noted that the judgment rule is that the prediction error exceeds the dynamic threshold, the model performance deteriorates significantly, or a new pattern is automatically detected in the data.

[0074] S4.6: Develop a comprehensive performance evaluation system to fully assess model performance by predicting interval width or coverage probability.

[0075] Specifically, an optimal prediction interval (PI) is dynamically generated for each predicted output using advanced statistical models or machine learning techniques; the prediction interval width (PIW) for each prediction is calculated using fuzzy logic, and the fuzzy average or fuzzy median is calculated to assess the model's prediction uncertainty; advanced probability theory methods are introduced to calculate the probability coverage (CP) of the predicted true value falling within the prediction interval; PIW and CP are analyzed in a unified manner, and a multi-objective optimization algorithm is used to comprehensively evaluate the overall performance of the model; an intelligent decision support system is introduced to dynamically determine the optimal prediction model configuration under different conditions based on the comprehensive evaluation results; a real-time performance monitoring system is built, with an embedded model self-optimization module that automatically tunes and optimizes the model when an increase in PIW or a decrease in CP is detected; the model is iteratively validated using the latest data, and continuously updated and optimized in conjunction with market demands and technological changes.

[0076] It should be noted that the specific formula for generating the optimal prediction interval PI is as follows:

[0077]

[0078] Where PI represents the optimal prediction interval. Z represents the predicted value. a denoted by quantile, SE represents the standard error of the normal distribution.

[0079] It should be noted that an ideal model should have a small PIW (indicating low prediction uncertainty) and a high CP (indicating high prediction accuracy).

[0080] S4.7: Establish a continuous model iteration and update mechanism based on performance evaluation results to dynamically optimize the model structure and parameters in order to adapt to environmental changes and technological advancements.

[0081] S5: Using a trained model, input real-time or latest collected weather forecast data and wind farm operation data to perform wind turbine output prediction.

[0082] S6: Evaluate the operational performance of the wind farm based on the forecast results to guide equipment maintenance and upgrades.

[0083] Furthermore, this embodiment also provides a wind turbine output prediction system based on meteorological intelligence, including a multi-source data collection module for collecting multi-source data including historical wind turbine output data, climate characteristic data, meteorological forecast data, and topographic data; a preprocessing module for preprocessing the collected data and automatically determining the resampling standard based on the time sequence pattern and trend of the data; a segmentation module for segmenting the dataset by comprehensively considering the time sequence, geographical and climatic characteristics of the data; a training module for training the model using LSTM and CNN fused with LSTM respectively to improve the ability to recognize long-term meteorological patterns and capture the spatial distribution characteristics of wind farms; a prediction module for using the trained model, inputting real-time or latest collected meteorological forecast data and wind farm operation status data, to predict wind turbine output; and an evaluation module for evaluating the wind farm operation effect based on the prediction results to guide equipment maintenance and renovation.

[0084] This embodiment also provides a computer device applicable to the wind turbine output prediction method based on meteorological intelligence, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the wind turbine output prediction method based on meteorological intelligence as proposed in the above embodiment.

[0085] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0086] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the wind turbine output prediction method based on meteorological intelligence as proposed in the above embodiments.

[0087] In summary, this invention achieves deep understanding and high-quality feature extraction of complex meteorological data by constructing an end-to-end intelligent meteorological system; it enables real-time and continuous forecasting of wind farms without human intervention by designing models with autonomous learning and adaptability; it evaluates operational performance based on forecast results, guiding equipment maintenance and upgrades, reducing costs, and improving economic efficiency; it achieves perception, analysis, decision-making, and control of the entire wind farm through end-to-end system integration of various intelligent algorithms; and it accelerates the replacement of traditional energy sources with wind power by improving the economic efficiency and reliability of wind power, thereby maximizing the environmental and social benefits of wind power.

[0088] Example 2

[0089] Reference Figures 1-2 This is the second embodiment of the present invention, which provides a wind turbine output prediction method based on meteorological intelligence. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiment.

[0090] Specifically, the largest wind farm A in a certain province was selected as the research object. Historical wind turbine output data, meteorological monitoring data, and wind farm operation data from the past three years were collected using the wind farm's existing monitoring system. Simultaneously, climate characteristic data for the region were obtained from the meteorological department, weather forecast data for the region for the next week were obtained from the numerical weather prediction system, and topographic data of the wind farm's location was extracted from a geographic information database.

[0091] Furthermore, tools such as Pandas and NumPy in Python were used to preprocess the collected data, including cleaning, anomaly removal, and normalization. Time series analysis tools such as FB Prophet were used to analyze data trends and periodic patterns, and appropriate resampling intervals were set.

[0092] Furthermore, based on data such as latitude, longitude, altitude, and vegetation cover of the wind farm's location, the dataset was split into training and testing datasets. The training dataset comprised 70% of the dataset, and the testing dataset comprised 30%. A hybrid network combining CNN and LSTM models was constructed using TensorFlow, with hyperparameters set as follows: learning rate 0.001, training epochs 100, time step 24, and the optimizer Adam. The model was then trained using the training data.

[0093] Furthermore, the trained model was validated using test data. The mean absolute error of the prediction results was 31.67%, which is nearly 12 percentage points higher than the 43.25% of the existing prediction model used in this wind farm.

[0094] Preferably, the comparative indicators of the present invention and the conventional method are shown in Table 1.

[0095] Table 1 Comparison Indicators between the Invention and Traditional Methods

[0096] Comparison indicators This invention Traditional methods Prediction accuracy 86.3％ 73.5％ Mean Absolute Error 31.6％ 43.25％ Iterative optimization efficiency (High) Monthly Update (Low) reliance on manual updates Resource utilization efficiency Increase by 40% No significant changes

[0097] Specifically, as shown in Table 1, considering the two most critical evaluation indicators—prediction accuracy and mean absolute error—this invention demonstrates a significant improvement in wind power output prediction accuracy, with an accuracy increase of 12.8 percentage points and a mean absolute error decrease of 11.58 percentage points. This showcases the outstanding effect of this invention in improving prediction quality. Furthermore, this invention establishes a continuous iteration mechanism, achieving automatic monthly optimization and demonstrating higher autonomous adaptability. In addition, in terms of computational resource utilization efficiency, this invention improves efficiency by 40% through parallel computing and other methods, reflecting better economic efficiency.

[0098] In summary, the technical solution of this invention demonstrates significant advantages in multiple aspects, including accuracy, efficiency, intelligence, and economy, in wind power output prediction. It fully showcases that, compared to traditional methods, the intelligent meteorological forecasting system constructed by this invention is more efficient, economical, and adaptable.

[0099] 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 scope of the claims of the present invention.

Claims

1. A wind turbine output prediction method based on meteorological intelligence, characterized in that: include, Collect multi-source data, including historical wind turbine output data, climate characteristic data, weather forecast data, and topographic data; The collected data is preprocessed, and the resampling criteria are automatically determined based on the time series pattern and trend of the data. The dataset is segmented by comprehensively considering the temporal order, geographical and climatic characteristics of the data; The model was trained using LSTM and CNN fused with LSTM respectively to improve the ability to identify long-term meteorological patterns and capture the spatial distribution characteristics of wind farms. Using a trained model, input real-time or latest collected weather forecast data and wind farm operation data, perform wind turbine output prediction; The forecast results are used to assess the operational effectiveness of wind farms in order to guide equipment maintenance and upgrades.

2. The wind turbine output prediction method based on meteorological intelligence as described in claim 1, characterized in that: The automatic determination of resampling criteria based on the temporal patterns and trends of data includes the following steps: Build a flexible and scalable data import module to support real-time acquisition of multi-source heterogeneous time series data; We use an unsupervised anomaly detection algorithm based on isolated forests to automatically identify low-quality data, and use data augmentation and generative models to automatically fill in missing values. Using an unsupervised binary clustering algorithm based on a tree model, the data distribution is evaluated in real time and the optimal normalization scheme is selected. The high-order time-dependent features of the data are learned through an automatic feature learning module based on deep neural networks. An autoregressive model based on Fourier and wavelet transform is used to evaluate data periodicity in real time and dynamically adjust the resampling interval. A Bayesian optimization search algorithm is used to search online for the optimal combination of preprocessing parameters and modules; By leveraging feedback from downstream prediction tasks, the preprocessing pipeline is fine-tuned end-to-end through multi-task learning. Using a locally interpretable model, explanations for decisions made at each preprocessing step are generated in real time; Online learning and incremental learning algorithms are used to enable the preprocessing model to continuously adapt to changes in data.

3. The wind turbine output prediction method based on meteorological intelligence as described in claim 1, characterized in that: The process of splitting the dataset includes the following steps: Graph convolutional networks are used to model potential relationships between regions, enriching the contextual representation of time-series data and capturing the spatiotemporal correlations between data. Construct a knowledge graph and integrate multi-source heterogeneous geographic environment data to perform external knowledge transfer and inject additional semantic expression into the original dataset; Construct a multi-temporal granularity integrated sub-model. If the time series exhibits a low-frequency periodic pattern, a coarser-grained model should be selected; if it reflects high-frequency data changes, a finer-grained model should be selected to comprehensively simulate various potential periodicities of the time series. We encapsulate a multi-agent adversarial search mechanism, formalize the dataset segmentation problem into a decision-making process, and use game theory to find robust segmentation strategies. Identify similar and comparable clusters in the dataset and apply transfer learning algorithms to perform adaptive knowledge transfer between clusters in order to improve the segmentation effect between them; Based on reinforcement learning theory, an agent environment is constructed and integrated feedback is continuously generated, enabling the segmentation model to evolve towards a better solution.

4. The wind turbine output prediction method based on meteorological intelligence as described in claim 1, characterized in that: The training of the model using LSTM and LSTM fused with CNN respectively includes the following steps: A comprehensive dataset is created by integrating multi-source data through a data fusion framework and using feature selection algorithms and generative feature construction. The LSTM model architecture was customized and improved to adapt to the temporal complexity of meteorological data; An improved CNN architecture was developed, an LSTM module was added, and multi-scale convolutional kernels were used to conduct in-depth analysis of the spatial distribution characteristics of wind farms. A fusion model is built by weighted fusion and model stacking to integrate the predictive capabilities of LSTM and CNN; Determine whether the fusion model needs fine-tuning based on real-time and historical performance data; Establish a comprehensive performance evaluation system to fully assess model performance by predicting interval width or coverage probability; A continuous model iteration and update mechanism is established based on performance evaluation results to dynamically optimize the model structure and parameters to adapt to environmental changes.

5. The wind turbine output prediction method based on meteorological intelligence as described in claim 4, characterized in that: The customized improvement of the LSTM model architecture includes the following steps: By increasing the number of layers in the LSTM network, a multi-layer LSTM structure is constructed to enhance the model's ability to capture and learn dependencies in long-term time-series data. In each LSTM layer, a dynamic neuron number configuration mechanism is implemented, and the optimal number of neurons in each layer is determined through experimental analysis. An attention mechanism is embedded in the LSTM structure to improve the LSTM network's ability to identify and weight key temporal features; A bidirectional LSTM structure is used to process the forward and reverse information flows of the sequence simultaneously, and to integrate past and future context information. Dropout and L1 / L2 regularization strategies are applied to the LSTM network, and overfitting of the LSTM model is prevented by randomly dropping some network connections and applying regularization penalties to the weights. Implement a dynamic learning rate adjustment mechanism, and automatically adjust the learning rate using an adaptive learning rate optimization algorithm; The customized and improved LSTM cell formula is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t = no t *fish(c t ) Where σ represents the Sigmoid function, f t i t and o t These represent the forget gate, input gate, and output gate, respectively. t Represents the cell state, and tanh represents the hyperbolic tangent activation function. The candidate cell state is represented by *, element-wise multiplication is represented by W and b, which represent the weight matrix and bias term, respectively. The subscripts f, i, c, and o correspond to the forget gate, input gate, cell state, and output gate, respectively. t h represents the input at a given time t. t-1 This represents the hidden state at the previous time step t-1.

6. The wind turbine output prediction method based on meteorological intelligence as described in claim 4, characterized in that: The development and improvement of the CNN architecture includes the following steps: Different sizes of convolutional kernels are used in parallel at the same level of the CNN model, enabling the model to capture both local details and global features in the wind farm simultaneously. By applying a spatial attention weighting mechanism to the convolutional feature maps, the CNN model resources are automatically identified and concentrated to focus on the key spatial features for wind power output prediction. By applying depthwise separable convolution techniques, the network complexity and computational resource consumption of CNN models can be reduced; Develop a dynamic convolutional kernel learning strategy that allows CNN models to adaptively adjust the size and shape of convolutional kernels based on specific wind farm layouts and environmental conditions; Implement feature pyramid networks or cross-layer connection techniques to fuse feature information from different levels.

7. The wind turbine output prediction method based on meteorological intelligence as described in claim 6, characterized in that: The adaptive adjustment of the size and shape of the convolution kernel according to the specific wind farm layout and environmental conditions includes, A real-time monitoring system is built by monitoring the data stream of the wind farm to determine the specific operating conditions of the wind farm. Construct decision rules, If the wind speed is consistently higher than the average wind speed and fluctuates little, a smaller convolutional kernel is selected to focus on capturing high-resolution local features in order to improve the accuracy of the prediction details. If a sudden change in wind speed is detected in a short period of time, the convolution kernel is dynamically adjusted to expand its shape and span multiple data points to capture sudden spatial variations. If the wind direction data shows regular periodic changes, then an L-shaped or J-shaped convolution kernel is used to lock in the pattern of periodic wind direction changes. If multiple turbulence superposition regions are detected in the wind farm, then cross-shaped or star-shaped convolution kernels are applied to focus on multiple intersection points to analyze the superposition of multiple modes. If the wind turbines in the wind farm are arranged in a non-uniform manner, resulting in extremely uneven local wind flow distribution, irregular or refined convolution kernels are used to capture the local wind flow characteristics in detail. The system captures various monitoring indicators of the wind farm in real time and performs preliminary analysis through a real-time monitoring system to assess the current environmental status. Based on data analyzed by a real-time monitoring system, decision rules are used to dynamically adjust the convolution kernel in order to capture key features; The adjusted CNN model predicts wind power output and compares it with actual measurements to calculate the prediction error. Analyze the sources of prediction errors to determine whether they are due to improper adjustment of the convolution kernel; Based on the error analysis results, adjust and optimize the parameters of the convolution kernel adjustment rule and the decision rule; When the system detects that the growth of prediction accuracy is slowing down or has reached a predetermined standard, the optimization loop is completed.

8. A wind turbine output prediction system based on meteorological intelligence, based on the wind turbine output prediction method based on meteorological intelligence as described in any one of claims 1 to 7, characterized in that: It also includes, The multi-source data collection module is used to collect multi-source data, including historical wind turbine output data, climate characteristic data, weather forecast data, and topographic data. The preprocessing module is used to preprocess the collected data and automatically determine the resampling standard based on the time series pattern and trend of the data. The segmentation module is used to segment the dataset by comprehensively considering the temporal order, geographical and climatic characteristics of the data; The training module is used to train the model using LSTM and an improved CNN respectively, in order to improve the ability to identify long-term meteorological patterns and capture the spatial distribution characteristics of wind farms. The prediction module is used to predict wind turbine output by taking into account the trained model and the latest collected weather forecast data and wind farm operation data. The assessment module is used to evaluate the operational performance of wind farms based on the prediction results, in order to guide equipment maintenance and upgrades.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the wind turbine output prediction method based on meteorological intelligence as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the wind turbine output prediction method based on meteorological intelligence as described in any one of claims 1 to 7.

Images