Land-based wind power base hoisting window period prediction method and system

Abstract

The embodiment of the present application provides a kind of land wind power base hoisting window period prediction method and system, belong to meteorological prediction technical field.The method comprises: collecting the meteorological data and topographic data of hoisting target area;Based on the meteorological data and the topographic data, the weather information prediction in future predetermined time is carried out;Based on the weather information prediction, the suitable hoisting window time corresponding to future predetermined time is determined;The suitable hoisting window time determined is output.The present application scheme realizes the accurate prediction of the strong convection electricity that influences hoisting operation, and outputs the corresponding suitable hoisting window period, which is referred to by workers, to ensure the safety of hoisting operation.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, specifically to a method and system for predicting the installation window period of onshore wind power bases. Background Technology

[0002] Wind energy is a clean and renewable energy source, and wind power has become an important direction for global renewable energy development. Due to the advantages of grid transmission conditions, low-wind-speed wind turbines in low-wind-speed areas have gradually become the mainstay of the wind power industry. These areas generally have lower average wind speeds and relatively more construction and hoisting windows. Because wind turbine equipment is large and requires hoisting operations, environmental requirements are high. Strong convective weather can easily affect hoisting operations and cause safety accidents. Therefore, determining suitable hoisting windows before wind turbine construction can effectively avoid the impact of severe weather on hoisting operations, which has significant positive implications for both the protection of wind power equipment and the prevention of safety accidents. Existing methods for determining suitable hoisting windows rely on subjective determination of suitable hoisting windows based on weather forecast information provided by conventional weather forecasts. This approach has a relatively limited scope and low accuracy, failing to meet the increasing hoisting needs of low-wind-speed turbines. To address the shortcomings of existing methods in determining suitable hoisting windows in terms of comprehensiveness and accuracy, a new method for predicting hoisting windows in onshore wind power bases is needed. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for predicting the hoisting window period for onshore wind power bases, so as to at least solve the problem that existing methods for determining suitable hoisting windows are insufficient in terms of comprehensiveness and accuracy.

[0004] To achieve the above objectives, the first aspect of the present invention provides a method for predicting the hoisting window period of an onshore wind power base. The method includes: collecting meteorological data and topographic data of the hoisting target area; predicting meteorological information within a predetermined future time based on the meteorological data and the topographic data; determining a suitable hoisting window time within the corresponding predetermined future time based on the predicted meteorological information; and outputting the determined suitable hoisting window time.

[0005] Optionally, the meteorological data includes: meteorological forecast data from a preset numerical model and historical meteorological data for the hoisting target area; wherein, the meteorological forecast data from the preset numerical model includes: global numerical weather forecast data and / or regional numerical weather forecast data.

[0006] Optionally, the meteorological information forecast includes: wind speed forecast, wind direction forecast, and lightning forecast.

[0007] Optionally, the wind speed prediction includes: downscaling the meteorological data to obtain wind speed prediction data for the hoisting target area; combining the wind speed prediction data and the terrain data for the hoisting target area based on a preset CNN network model to simulate the wind field distribution in the hoisting target area; and predicting the wind speed within a predetermined time period based on the simulated wind field distribution and a preset LSTM model to obtain the wind speed prediction result.

[0008] Optionally, the method further includes: wind speed prediction result correction, including: obtaining the wheel hub height of the wind turbine to be installed; correcting the wind speed prediction result based on the wheel hub height to obtain a wind speed prediction result corresponding to the wheel hub height; wherein, the correction function for the wind speed prediction result is:

[0009]

[0010] in, Wind speed information at wheel height; The wind speed information in the wind speed prediction results is shown in z; z represents the wheel hub height; z b denoted as the height of the wind speed prediction result; 'a' is the surface roughness coefficient of the target hoisting area.

[0011] Optionally, the wind direction prediction includes: predicting meteorological data based on a preset EEMDSE-ILSTM combined prediction model to obtain wind direction prediction results; and processing the wind direction prediction results based on a preset wind direction change prediction model to obtain wind direction prediction results within a predetermined time period in the future.

[0012] Optionally, the lightning prediction includes: acquiring lightning observation data; wherein the lightning observation data includes: the time of lightning occurrence, latitude and longitude, current intensity, lightning type and steepness; preprocessing the lightning observation data; wherein the preprocessing rules include: data quality control, data class imbalance handling, data temporal and spatial alignment, feature engineering processing and feature selection processing; and training the preprocessed lightning observation data based on a preset ST-LSTM algorithm to obtain lightning prediction data for the hoisting target area.

[0013] Optionally, the determination of a suitable hoisting window time within a predetermined future time period based on predicted meteorological information includes: comparing wind speed prediction results with preset suitable hoisting wind speeds to determine a first suitable hoisting period; comparing wind direction prediction results with preset suitable hoisting wind directions to determine a second suitable hoisting period; determining a third suitable hoisting period based on lightning prediction results; comparing the first, second, and third periods to extract overlapping periods and forming a candidate set of suitable hoisting periods; and selecting suitable hoisting periods longer than the preset hoisting demand time from the candidate set of suitable hoisting periods as suitable hoisting window times based on the preset hoisting demand time.

[0014] A second aspect of the present invention provides a system for predicting the hoisting window period for onshore wind power bases. The system includes: a data acquisition unit for acquiring meteorological data and topographic data of the hoisting target area; a processing unit for predicting meteorological information within a predetermined future time based on the meteorological data and the topographic data; a window time determination unit for determining a suitable hoisting window time within the corresponding predetermined future time based on the predicted meteorological information; and an output unit for outputting the determined suitable hoisting window time.

[0015] On the other hand, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described method for predicting the hoisting window period of onshore wind power bases.

[0016] Through the aforementioned technical solutions, high-resolution regional numerical weather prediction model data, combined with multi-year stratified wind speed observation data from existing wind measurement towers and forecast correction technology based on deep learning algorithms, are used to develop accurate wind speed and direction predictions for wind farm locations and individual wind turbine locations. Simultaneously, using real-time lightning monitoring data and high-resolution satellite cloud imagery data, based on deep learning algorithms, accurate predictions of future lightning strike areas are achieved. Based on the actual needs of safe construction during wind farm installation, by comprehensively judging the actual wind speed and direction, predicted wind speed and direction, and lightning strike area prediction results, the system provides alerts and warnings of severe wind and lightning strike risks. It also achieves accurate prediction of strong convective electrical currents affecting installation operations, outputting corresponding suitable installation windows for workers' reference, thus ensuring the safety of installation operations.

[0017] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a flowchart of the steps of a method for predicting the hoisting window period of an onshore wind power base according to one embodiment of the present invention;

[0020] Figure 2 This is a flowchart of the meteorological information prediction steps provided by one embodiment of the present invention;

[0021] Figure 3 This is a flowchart of the steps for wind speed prediction provided by one embodiment of the present invention;

[0022] Figure 4 This is a flowchart of the lightning prediction steps provided by one embodiment of the present invention;

[0023] Figure 5 This is a system structure diagram of an onshore wind power base hoisting window prediction system provided in one embodiment of the present invention. Detailed Implementation

[0024] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0025] Wind energy is a clean and renewable energy source, and wind power has become an important direction for the development of renewable energy globally. Due to the advantages of grid transmission conditions in low-wind-speed areas, low-wind-speed wind turbines have gradually become the mainstay of the wind power industry. These areas generally have lower average wind speeds, resulting in more construction and installation windows. Because wind turbine equipment is large and requires hoisting operations, environmental requirements are high. Strong convective weather can easily affect hoisting operations and cause safety accidents. Therefore, determining suitable hoisting windows before wind turbine construction can effectively avoid the impact of severe weather on hoisting operations, which has significant positive implications for both the protection of wind power equipment and the prevention of safety accidents.

[0026] Wind is a key element in conventional weather forecasting. However, meteorological bureaus only provide wind speed levels over a wide area in their forecasts. Due to the technical limitations of numerical weather prediction, wind speed forecasts from numerical models are grid-averaged, which is insufficient in predicting maximum wind speeds and reflecting the local characteristics of wind speed. Furthermore, because wind speed is highly localized and abrupt, traditional wind forecast correction models often lack extensive historical observation data, making it impossible to make targeted corrections to numerical model wind speed predictions. Therefore, wind forecasts in conventional weather forecasts often fail to meet the needs of practical engineering projects. In addition, lightning is also a potential threat to the safety of wind turbine installation, and current conventional weather forecast services often lack monitoring and forecasting of lightning.

[0027] Existing methods rely on subjective weather forecasts to determine suitable hoisting windows, which are limited in scope and have low accuracy, failing to meet the increasing hoisting demands of low-wind-speed turbines. Therefore, this invention proposes a method for predicting hoisting windows at onshore wind farms. It utilizes high-resolution regional numerical weather prediction model data, combined with multi-year stratified wind speed observation data from existing anemometers and forecast correction technology based on deep learning algorithms, to develop accurate wind speed and direction predictions for the wind farm's location and each turbine's position. Simultaneously, it uses real-time lightning monitoring data and high-resolution satellite cloud imagery, based on deep learning algorithms, to achieve accurate predictions of future lightning strike areas. Based on the actual needs of safe construction during wind farm hoisting, it comprehensively judges the actual wind speed and direction, predicted wind speed and direction, and lightning strike area prediction results to provide alerts and warnings of severe wind and lightning risks. It achieves accurate prediction of strong convective electrical forces affecting hoisting operations, outputting corresponding suitable hoisting windows for operators' reference, thus ensuring the safety of hoisting operations.

[0028] Figure 2 This is a system structure diagram of an onshore wind power base hoisting window prediction system provided in one embodiment of the present invention. Figure 2 As shown, this invention provides a system for predicting the hoisting window period for onshore wind power bases. The system includes: a data acquisition unit for acquiring meteorological and topographic data of the hoisting target area; a processing unit for predicting meteorological information within a predetermined future time based on the meteorological and topographic data; a window time determination unit for determining a suitable hoisting window time within the corresponding predetermined future time based on the predicted meteorological information; and an output unit for outputting the determined suitable hoisting window time.

[0029] Figure 1 This is a flowchart of a method for predicting the installation window period of onshore wind power bases according to one embodiment of the present invention. Figure 1 As shown, this invention provides a method for predicting the installation window period for onshore wind power bases, the method comprising:

[0030] Step S10: Collect meteorological and topographic data of the target hoisting area.

[0031] Specifically, meteorological forecast data from preset numerical models and historical meteorological data of the target area are collected through preset ports. The meteorological forecast data includes global numerical weather prediction data and / or regional numerical weather prediction data. Numerical weather prediction refers to a method that, based on actual atmospheric conditions and certain initial and boundary conditions, uses large-scale computers to perform numerical calculations to solve fluid dynamics and thermodynamic equations describing weather evolution, predicting atmospheric motion and weather phenomena over a future period. This invention uses numerical weather prediction data as the foundational data, and then uses innovative scaling and correction algorithms to accurately predict weather conditions in the target area. Because numerical weather prediction data is open data, it can be directly collected through preset network ports. This data includes systems such as ECMWF (European Centre for Medium-Range Weather Forecasts), GFS (United States Weather Forecast Center), GEM (Germany Weather Forecasting Centre), ICON (Germany Weather Forecasting Centre), and GRAPES (China Regional General Numerical Weather Prediction), etc. In theory, any open database that can provide numerical weather forecast data can be used as the basis for this application.

[0032] Historical meteorological data for the target hoisting area can be extracted from historical monitoring data from local meteorological stations or from local meteorological forecast data in open databases.

[0033] As for the terrain data of the hoisting area, it includes the altitude, topographic information, and functional information of the target area. For example, it determines the terrain information of the target area as an island, coast, lake shore, desert, field, countryside, jungle, hill, or urban area, which is used for subsequent weather analysis in conjunction with the terrain.

[0034] Step S20: Based on the meteorological data and the terrain data, predict the meteorological information for a predetermined period of time in the future.

[0035] Specifically, meteorological forecasting includes wind speed forecasting, wind direction forecasting, and lightning forecasting. These three parts need to be forecasted separately, so specifically, for example... Figure 2 This includes the following steps:

[0036] Step S201: Perform wind speed prediction.

[0037] Specifically, to build wind speed prediction models for each wind turbine location in a wind farm, accurate meteorological element forecasts for each location are needed. Since the resolution of existing numerical weather prediction models is generally around 10km, they cannot effectively predict small-scale weather phenomena. Therefore, downscaling techniques are needed to improve the accuracy of numerical weather prediction models and obtain meteorological forecast data for each wind turbine location. Traditional downscaling methods generally include three types: dynamic downscaling, statistical downscaling, and a combination of dynamic and statistical downscaling. Dynamic downscaling uses large-scale data as input and boundary conditions for small-scale models, essentially using large-scale data as the initial and boundary conditions for high-precision atmospheric models. However, increasing model resolution and improving various aspects of the dynamic model itself, as well as refining the model's physical processes and parameterization schemes, to improve forecast accuracy is very difficult. Dynamic downscaling also suffers from high computational costs. Statistical downscaling, on the other hand, establishes element forecasting models based on the statistical relationships between model outputs and meteorological elements, making quantitative forecasts for these elements. The basic principle of this method is to use statistical empirical methods to establish linear or nonlinear statistical relationships between forecast factors and elements using observational data, and then test these relationships using independent observational data. Statistical downscaling methods are relatively flexible and computationally inexpensive, so they are frequently used in climate assessments. Currently, the main statistical downscaling methods applicable to numerical products include Model Output Statistics (MOS), Full Prediction (PP), Kalman filtering, and deep learning methods.

[0038] To address the limitations of traditional downscaling methods, this invention proposes a downscaling technique based on artificial intelligence algorithms. Utilizing low-resolution model forecast data and high-resolution topographic information, the LapSRN algorithm is employed to obtain refined meteorological element forecasts, achieving more efficient and accurate downscaling. LapSRN performs different levels of convolution and upsampling on the initial image. The upsampled feature map is then used to obtain sub-band residual information, which is added to the upsampled medium-resolution image to obtain a medium-resolution image. This process is repeated incrementally to obtain higher-resolution images. Due to its unique internal filter structure, weight sharing, and translation invariance, it is widely used in image processing. When improving the spatial resolution of the data, considering regional and topographic factors, the model selects the original data and topographic information as input, using a convolutional neural network to extract high-dimensional features for high-resolution data reconstruction.

[0039] The first stage of the LapSRN model takes low-spatial-resolution model data and high-spatial-resolution terrain information as input to obtain high-spatial-resolution data. The second stage takes the obtained high-spatial-resolution data and even higher-spatial-resolution terrain information as input to obtain even higher-spatial-resolution data, and so on. After scaling the meteorological data, wind speed prediction data for the target area is obtained.

[0040] To achieve accurate wind speed forecasting, this invention proposes two wind speed forecasting modes: ultra-short-term forecasting and short-term forecasting. Ultra-short-term forecasting provides precise wind speed predictions every 15 minutes for at least the next 4 hours, while short-term forecasting provides precise wind speed predictions every 15 minutes for at least the next 3 days. Specifically, as... Figure 3 This includes the following steps.

[0041] Step S2011: Conduct ultra-short-term wind speed forecasting.

[0042] For ultra-short-term forecasting, when modeling wind for the ultra-short term, due to the continuity of wind speed and direction data, in addition to accurate meteorological forecast data for the location of wind turbines, recent actual wind data is also needed as input to the model. Wind speed curves are volatile, and directly using time-series data as input is not conducive to model learning and feature extraction. Therefore, signal decomposition techniques can be used to decompose the actual wind speed time-series data, extract trend features, obtain relatively stable components, and then filter each component using sample entropy to simplify the data. The decomposed signals and other relevant meteorological element data are then used as features and input into the ILSTM model. The prediction results of each component are output and accumulated to obtain the final predicted value.

[0043] Empirical Mode Decomposition (EMD) is a method for stabilizing nonlinear and non-stationary signals. EMD decomposes complex nonlinear signals into a series of smooth IMF components. These IMF components contain local features of the original signal's fluctuations and trends at different scales, which helps in understanding the signal's true physical meaning. EMD is considered a major breakthrough in linear and steady-state spectral analysis based on Fourier transforms since 2000. This method decomposes signals based on the data's own time-scale characteristics without requiring pre-defined basis functions, a fundamental difference from Fourier and wavelet decomposition methods based on a priori harmonic and wavelet basis functions. Due to this characteristic, EMD can theoretically be applied to any type of signal decomposition, thus offering significant advantages in processing non-stationary and nonlinear data. It is suitable for analyzing nonlinear and non-stationary signal sequences and boasts a high signal-to-noise ratio. However, EMD can produce mode aliasing, leading to a loss of the physical meaning of the IMFs and a decrease in the adaptability of the prediction model. The Ensemble Empirical Mode Decomposition (EEMD) algorithm can overcome the shortcomings of the EMD algorithm. The EEMD algorithm uses noise-assisted signal processing to add white noise to the original signal and solves the mode mixing problem by averaging the multiple IMFs of the decomposition. Sample entropy (SE) is a time series complexity test method, mainly used to quantitatively describe the complexity and regularity of a system. Compared with approximate entropy, sample entropy has the advantages of not calculating self-matching degree and having high tolerance for missing data, thus overcoming the bias of approximate entropy calculation and having a faster calculation speed.

[0044] Long Short-Term Memory (LSTM) is a special type of RNN, primarily designed to address the vanishing and exploding gradient problems during training of long sequences. Simply put, compared to ordinary RNNs, LSTM performs better with longer sequences.

[0045] The difference between LSTM and RNN lies in the fact that LSTM adds an information conveyor belt at the top layer of the network to store information. An LSTM network has four neural network layers: a forgetting layer, an input layer, an update layer, and an output layer. Its workflow can be summarized as follows:

[0046] First, the forgetting layer controls the information that can pass through using an activation function (sigmoid), utilizing the output h from the previous time step. t-1 and the current input x t To generate a forgetting f between 0 and 1 tValue, according to f t The value determines whether to let the information c from the previous moment be stored. t-1 Passed. The calculation formula is:

[0047] f t =σ(W f *[h t-1 ,x t ]+b f )

[0048] Then, new information that needs to be updated is generated. First, the input layer determines which value needs to be updated through an activation function, and then the tanh function generates new candidate values. Add it to the state. The calculation formula is:

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

[0050]

[0051] Among them, i t The input gate represents the current state of the input unit, W. i and W c b is the weight value. i and b c This is a bias term.

[0052] Furthermore, update the old state, and put c t-1 The status is updated to c t c t-1 The unnecessary information is forgotten through the forgetting gate, and a new candidate value is added to arrive at a new state. The calculation formula is:

[0053]

[0054] Among them, c t In cellular state, This indicates point-by-point multiplication.

[0055] Finally, the model's output is determined. First, an initial output is generated using an activation function, then the tanh function is used to adjust c. t The values ​​are scaled to (-1, 1) and then multiplied pairwise with the output of the activation function to obtain the model's output. The calculation formula is:

[0056] o t =σ(W o *[h t-1 ,x t ]+b o )

[0057]

[0058] In the formula, o t For output gate, h t W is the output value at time t. o b is the output gate weight value. o This is used to bias the output gate.

[0059] Step S2012: Conduct short-term wind speed forecasting.

[0060] Specifically, when forecasting short-term wind speeds at wind turbine locations in wind power bases, a CNN prediction model is selected. Using accurate meteorological forecast data for each turbine location obtained through downscaling, combined with time and geographic information, appropriate data is selected as features through feature engineering and feature selection. These features are then input into the CNN model in chronological order, ultimately outputting wind speed and direction predictions for at least three days, every 15 minutes, thus achieving short-term wind forecasting.

[0061] CNNs (Neural Networks) are multi-layered neural networks that excel at handling image processing, especially large image processing problems. CNNs successfully reduce the dimensionality of massive image recognition datasets through a series of methods, making them trainable. The difference between CNNs and ordinary neural networks lies in the fact that CNNs contain multiple feature extractors composed of convolutional and pooling layers. In a CNN's convolutional layer, a neuron is connected only to a subset of its neighboring neurons. A CNN convolutional layer typically contains several feature maps, each composed of neurons arranged in a rectangular pattern. Neurons within the same feature map share weights, which are called the convolutional kernel. The convolutional kernel is usually initialized as a random fractional matrix, and during network training, it learns appropriate weights. The direct benefit of shared weights (convolutional kernels) is reduced connections between network layers, thus lowering the risk of overfitting. Subsampling, also called pooling, typically takes two forms: mean pooling and max pooling. Subsampling can be viewed as a special type of convolution process.

[0062] For short-term wind speed forecasting at wind turbine locations in wind power bases, a convolutional neural network (CNN) model is selected for prediction. When forecasting wind speed and direction every 15 minutes for at least the next 3 days, wind speed and direction data, along with time and geographical information, are extracted from numerical model forecasts, arranged in chronological order, and input into the model. Finally, the wind speed and direction forecast results for each wind turbine location for at least the next 3 days are output every 15 minutes.

[0063] Preferably, after completing the wind speed prediction, wind speed correction is also required. To meet construction needs, wind speed and direction forecasts at specific heights are required. Therefore, wind speed and direction can be corrected based on the aforementioned 10-meter, 100-meter ultra-short-term, and short-term wind forecasts. Specific steps: First, the wind speed at the required height is initially calculated using the wind speed-height conversion formula; then, the wind speed correction model is trained using historical real-time wind data from different heights of the meteorological tower; finally, the wind speed data processed by the wind speed conversion formula is used as the input to the correction model DNN, and the output is the required wind speed prediction data at the specific height. The correction function for the wind speed prediction result is:

[0064]

[0065] in, Wind speed information at wheel height; This refers to the wind speed information in the wind speed prediction results;

[0066] z is the wheel height; b 'a' represents the height of the predicted wind speed; 'a' represents the surface roughness coefficient of the target hoisting area. Different roughness categories are categorized as shown in Table 1:

[0067]

[0068] Table 1 Roughness coefficients for different terrains

[0069] Deep Neural Networks (DNNs) can be understood as a type of neural network with many hidden layers, also known as a Multilayer Perceptron (MLP). Based on their position, DNN layers can be divided into three categories: input layers, hidden layers, and output layers. Generally, the first layer is the input layer, the last layer is the output layer, and the layers in between are hidden layers. DNN layers are fully connected, meaning that any neuron in the i-th layer is connected to any neuron in the (i+1)-th layer. Although DNNs appear complex, from a small local model perspective, they are similar to a perceptron, consisting of a linear relationship and an activation function.

[0070] Step S202: Perform wind direction prediction.

[0071] Specifically, similar to the wind speed prediction in step S201, it is also necessary to predict the wind direction based on the scaled prediction information. The only difference is that the final output is the predicted wind direction. The implementation principles can be interchanged, so they will not be elaborated here.

[0072] Step S203: Perform lightning prediction.

[0073] Specifically, due to their height, the locations where wind turbines are installed are susceptible to lightning strikes, making lightning strike zone prediction necessary to ensure construction safety. Traditional prediction methods mainly include the following two:

[0074] 1): Short-term prediction based on real-time lightning monitoring data, with real-time detection data as the main source and radar and satellite data as supplementary sources, to predict the lightning strike area;

[0075] 2) Short-term forecasting based on weather radar monitoring data and high-resolution satellite cloud imagery data involves extrapolating the images to obtain future radar and satellite data, thus enabling lightning prediction. Traditional forecasting methods are built upon extensive theoretical research and practical experience. The characteristics of each variable have relatively clear physical meanings, and parameter and threshold settings utilize both theoretical derivations and statistical analysis methods, as well as empirical formulas based on large amounts of observational data. Artificial intelligence algorithms, through error backpropagation mechanisms, can automatically search for optimal points in the parameter space and can easily handle multi-source heterogeneous data. This means they can effectively utilize various other observational data and automatically extract features from the input data through variable space transformation or deep neural networks. By appropriately combining traditional algorithms and artificial intelligence technologies, their respective advantages can be fully utilized, compensating for each other's weaknesses and complementing each other, thereby improving the accuracy of short-term lightning forecasts.

[0076] The short-term prediction model based on the combination of artificial intelligence and traditional algorithms mainly utilizes data such as radar, satellite, ground and satellite lightning location observation, ground meteorological observation, radar extrapolation, and lightning strike area extrapolation results based on lightning data. It uses artificial intelligence algorithms to directly predict the probability distribution of lightning occurrence, and uses areas with a probability greater than a certain threshold as lightning strike areas and provides graded warnings, so as to achieve accurate prediction of lightning strike areas at least 2 hours in 10-minute increments.

[0077] Specifically, the meteorological data is preprocessed; the preprocessing rules include: data quality control, data resolution reconstruction, feature engineering, and feature selection. The preprocessed meteorological data is then used to predict wind direction based on a pre-set EEMDSE-ILSTM combined prediction model to obtain wind direction prediction results. Finally, the wind direction prediction results are processed based on a pre-set wind direction change prediction model to obtain wind direction prediction results for a predetermined future time period. Specifically, as... Figure 4 Data preprocessing includes the following steps:

[0078] Step S2031: Data quality control.

[0079] Specifically, the collected data first undergoes necessary data quality control to eliminate radar clutter, satellite image stripes, outliers, suppress noise, and fill in missing values. Radar clutter needs to be removed according to its type, while other noise is smoothed using a median filter, and missing values ​​are filled using bilinear interpolation.

[0080] Step S2032: Handling data class imbalance.

[0081] Specifically, lightning is an extreme weather event with a low probability of occurrence in real life. Data class imbalance is a common problem in classification tasks. If this problem is not addressed, the model will tend to predict the class with the larger sample size, ignoring the class with the smaller sample size. This can lead to significant losses in practical applications due to missed classifications and other errors.

[0082] Common methods for addressing class imbalance include upsampling, downsampling, and loss-sensitive learning. Upsampling, for example, uses methods like sampling with replacement or SMOTE to increase the amount of data for the class with less data. The basic idea of ​​the SMOTE algorithm is to analyze and simulate minority class samples and add these artificially simulated new samples to the dataset, thus correcting the severe class imbalance in the original data. The simulation process uses the KNN technique, and the steps for generating new samples are as follows: First, the nearest neighbor algorithm is used to calculate the K nearest neighbors for each minority class sample. Then, N samples are randomly selected from these K nearest neighbors and subjected to random linear interpolation to construct new minority class samples. Finally, the new samples are combined with the original data to generate a new training set, achieving class balance.

[0083] Step S2033: Data temporal and spatial alignment.

[0084] Specifically, due to the differences in spatiotemporal resolution among various observation data, high spatial resolution reconstruction and temporal-spatial dimension matching and alignment are required. Because of the increased data sources and the large number of input features, bilinear interpolation is used for refinement and alignment of all channels. When labeling using historical real-time data, a relatively fine (1km) grid can be generated. When labeling data, traditional methods are used to comprehensively consider data such as satellite-observed radiation brightness temperature (TBB) data, cloud echo attributes calculated using multi-elevation angle basic reflectivity factors, and ground and satellite lightning location observations to delineate lightning strike areas. In areas with radar coverage, radar echoes are used as the primary source, while in areas without radar coverage, satellite observation data is used as the primary source.

[0085] Step S2034: Feature engineering and feature selection processing.

[0086] Specifically, although neural networks can theoretically extract and select features automatically, feature engineering and feature selection based on traditional models and statistical analysis can effectively reduce model complexity and training difficulty.

[0087] Feature subset selection is part of feature engineering. This is because among multiple features, there may be several that are unrelated to the response variable. Introducing too many features not only greatly increases the computational load but also affects the evaluation of the model. Therefore, we need to select variables that are related to the response variable to form a feature subset.

[0088] Optimal Subset: Starting with model zero (null model) M0, which only has an intercept term and no independent variables, we fit different feature combinations and select the best model (minimum RSS or maximum R²) from each. This means model M1 with one feature, model M2 with two features, and so on, up to model Mp with p features. Then, we select the best model from these p+1 models (based on cross-validation error, Cp, BIC, or adjusted R²). The features configured in this best model are the selected features. Because all possible feature combinations have been explored, the selected features are guaranteed to be optimal.

[0089] After data preprocessing, the preprocessed lightning observation data is trained using a pre-defined ST-LSTM algorithm to obtain lightning prediction data for the hoisting target area. The pre-defined ST-LSTM algorithm is an improvement on ConvLSTM based on PredRNN, allowing cross-layer connections between units belonging to different layers. It stores spatial and temporal features in a unified memory unit and transfers memory at both the vertical and horizontal levels. The underlying idea of ​​PredRNN is that a predictive learning system should simultaneously store spatial representations and temporal changes in a unified memory pool. The figure below illustrates the structural differences between stacked ConvLSTM and PredRNN.

[0090] In a typical stacked ConvLSTM, features are simply abstracted upwards step by step between layers, with cell states only propagated horizontally. This means that while there is memory flow between different time steps within the same layer, units in different layers at the same time step are independent and do not have memory flow. Therefore, spatial information is only propagated upwards along the hidden state. If the spatial information at the bottom layer at time t is particularly important, it cannot be propagated to the top layer at time t because vertically independent units lack memory.

[0091] This temporal memory flow is reasonable in spatiotemporally supervised learning because, based on research on stacked convolutional layers, the hidden representations from the bottom up can become increasingly abstract and class-specific. However, in spatiotemporally predictive learning, detailed information from the original input sequence should be preserved. To predict the future, features extracted from different levels of convolutional layers need to be learned. Therefore, a simple improvement is the structure of Deep Transition ConvLSTMs. The spatiotemporal memory units are updated in a zigzag direction, with information first passed up layers and then forward over time. However, this structure still has drawbacks: removing the horizontal temporal flow sacrifices temporal consistency because there is no temporal flow between different times within the same layer; the memory needs to travel longer paths between distant states, making it more prone to gradient vanishing.

[0092] PredRNN improves upon the traditional LSTM internal structure by introducing a new building block called ST-LSTM, which allows memory flow to occur simultaneously in both horizontal and vertical directions. The ST-LSTM unit structure is shown in the figure below. Unlike simple memory splicing, the ST-LSTM unit uses a shared output gate for both memory types, achieving seamless memory fusion and effectively modeling shape deformation and motion trajectories in spatiotemporal sequences.

[0093] Radar datasets accumulate, dissipate, or change slowly or rapidly due to various weather conditions, and also contain significant noise caused by factors such as terrain, making prediction extremely difficult. By comparing PredRNN with ConvLSTM and VPN baselines, PredRNN performs significantly better than ConvLSTM and VPN baselines. However, PredRNN still suffers from the gradient vanishing problem; through backpropagation over time, the gradient magnitude decays exponentially, which is essentially a problem of gradient vanishing due to long-term dependencies and training. PredRNN++ is an improvement on PredRNN, mainly in two aspects: a Casual LSTM unit was designed to replace the original ST-LSTM unit; and a Gradient Highway Unit (GHU) was designed to prevent gradient vanishing caused by long time series.

[0094] Step S30: Determine the appropriate hoisting window time within the corresponding future scheduled time based on the predicted meteorological information.

[0095] Specifically, by comparing the wind speed forecast results with the preset suitable hoisting wind speed, a first suitable hoisting time period is determined; by comparing the wind direction forecast results with the preset suitable hoisting wind direction, a second suitable hoisting time period is determined; by comparing the lightning forecast results, a third suitable hoisting time period is determined; by comparing the first, second, and third time periods, all overlapping time periods are extracted to form a candidate set of suitable hoisting time periods; based on the preset hoisting requirement time, suitable hoisting time periods longer than the preset hoisting requirement time are selected from the candidate set of suitable hoisting time periods as suitable hoisting window times.

[0096] Step S40: Output the appropriate hoisting window time.

[0097] Specifically, once a suitable lifting window is obtained, the result can be pushed to the user's terminal and displayed through a preset display unit, informing the user of the future suitable lifting time window for selection. This ensures the safety of the lifting operation.

[0098] The present invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described method for predicting the hoisting window period of onshore wind power bases.

[0099] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0100] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.

[0101] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.

Claims

1. A method for predicting the installation window period at onshore wind power bases, characterized in that, The method includes: Collect meteorological and topographic data of the target hoisting area; Based on the meteorological data and the topographic data, meteorological information is predicted for a predetermined period in the future; wherein... The meteorological information forecasts include: wind speed forecast, wind direction forecast, and lightning forecast; The wind speed prediction includes: downscaling the meteorological data to obtain wind speed prediction data for the hoisting target area; combining the wind speed prediction data and the terrain data for the hoisting target area based on a preset CNN network model to simulate the wind field distribution in the hoisting target area; and predicting the wind speed within a predetermined time period based on the simulated wind field distribution and a preset LSTM model to obtain the wind speed prediction result. Using low-resolution model forecast data and high-resolution terrain information, the LapSRN algorithm is employed to obtain refined meteorological element forecast results. The first stage of the LapSRN model takes low spatial resolution model data and high spatial resolution terrain information as input to obtain high spatial resolution data. The second stage takes the obtained high spatial resolution data and even higher spatial resolution terrain information as input to obtain even higher spatial resolution data. This process continues until the meteorological data is scaled up to obtain wind speed prediction data for the hoisting target area. The wind speed prediction results are corrected, including: initially calculating the wind speed at the required height level using the wind speed-height conversion formula; training a wind speed correction model using historical real-time wind data at different height levels of the meteorological tower; using the wind speed data processed by the wind speed conversion formula as input to the correction model DNN, and outputting the required wind speed prediction data at the specific height level; wherein, the correction function for the wind speed prediction results is: in, Wind speed information at wheel height; This refers to the wind speed information in the wind speed prediction results; This refers to the wheel height; denoted as the height of the wind speed prediction result; 'a' is the surface roughness coefficient of the target hoisting area. The lightning prediction process includes: acquiring lightning observation data; wherein the lightning observation data includes: the time of lightning occurrence, latitude and longitude, current intensity, lightning type, and steepness; preprocessing the lightning observation data; wherein the preprocessing rules include: data quality control, data class imbalance handling, data temporal and spatial alignment, feature engineering, and feature selection; and training the preprocessed lightning observation data based on a preset ST-LSTM algorithm to obtain lightning prediction data for the hoisting target area. Based on the predicted meteorological information, the appropriate hoisting window time for the corresponding future scheduled time is determined; Output the determined appropriate hoisting window time.

2. The method according to claim 1, characterized in that, The meteorological data includes: Meteorological forecast data from a preset numerical model and historical meteorological data for the target hoisting area; among which, The meteorological forecast data of the preset numerical model includes: Global numerical weather prediction data and / or regional numerical weather prediction data.

3. The method according to claim 1, characterized in that, The wind direction prediction includes: The meteorological data is predicted based on the pre-set EEMDSE-ILSTM combined prediction model to obtain wind direction prediction results. The wind direction prediction results are processed based on a preset wind direction change prediction model to obtain wind direction prediction results for a predetermined time period in the future.

4. The method according to claim 1, characterized in that, The determination of the appropriate hoisting window time within a corresponding future predetermined time period based on predicted meteorological information includes: By comparing the wind speed forecast results with the preset suitable wind speed for hoisting, the first suitable time period for hoisting is determined. By comparing the wind direction forecast results with the preset suitable wind direction for hoisting, a second suitable time period for hoisting is determined; The third suitable time period for hoisting was determined based on the lightning forecast results; By comparing the first time period, the second time period, and the third time period, the overlapping time periods of the three time periods are extracted to form a candidate set of suitable hoisting time periods; Based on the preset hoisting requirement time, suitable hoisting periods longer than the preset hoisting requirement time are selected from the candidate set of suitable hoisting periods and used as suitable hoisting window times.

5. A prediction system for the installation window period of an onshore wind power base, characterized in that, The system is used to execute the onshore wind power base hoisting window prediction method according to any one of claims 1-4, the system comprising: The data acquisition unit is used to collect meteorological and topographic data of the target hoisting area; The processing unit is used to predict meteorological information within a predetermined time period based on the meteorological data and the terrain data. The window time determination unit is used to determine the appropriate hoisting window time within the corresponding future predetermined time based on predicted meteorological information. The output unit is used to output the determined suitable hoisting window time.

6. A computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the onshore wind power base hoisting window prediction method according to any one of claims 1-4.

Images