Hybrid neural network-based joint prediction method and device for water and wind power generation power

Abstract

The present disclosure relates to the technical field of power generation prediction, and particularly relates to a water-wind-solar power generation power joint prediction method and device based on a hybrid neural network. A multi-source meteorological observation dataset of a target area is acquired; a historical power generation power dataset of a water-wind-solar power station in the target area is acquired; a GRU-CNN water-wind-solar power generation power joint prediction model is established and trained according to the multi-source meteorological observation dataset and the historical power generation power dataset; meteorological forecast data of the target area in a to-be-predicted time period is acquired in real time and transmitted to the trained GRU-CNN water-wind-solar power generation power joint prediction model to predict power generation power. The present disclosure improves the initial field through data assimilation to improve the numerical prediction effect, fuses the space-time correlation and load curve of water-wind-solar power generation power, constructs an improved GRU-CNN water-wind-solar power generation power joint prediction model, realizes water-wind-solar power generation power joint prediction, and improves the accuracy of water-wind-solar power generation power prediction.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of power generation prediction, in particular to a water, wind and light power generation joint prediction method and device based on a hybrid neural network. BACKGROUND

[0002] Developing clean energy such as water, wind and light is a major strategic measure to ensure future energy security and combat global climate warming. Wind energy and solar energy are the most promising new energy sources for large-scale development. However, both are easily affected by weather factors, and their output power has strong random fluctuations and is difficult to predict. With the continuous advancement of wind and light power generation base construction, large-scale wind and light power directly connected to the grid will bring great pressure to the peak, peak regulation and stable operation of the power system. Water turbine units have the characteristics of rapid start and stop, flexible operation, large output change amplitude and fast response to load changes, and are ideal peak regulation power sources. By taking advantage of the natural complementarity of resources and the flexibility of water power, multiple energy sources such as water, wind and light can be aggregated to form a multi-energy complementary power generation system, which is an effective way to reduce the impact of new energy grid connection and improve the pass rate of basin resources.

[0003] Current power prediction methods are all for power prediction of water power, wind power and photovoltaic power generation, without considering the influence between water, wind and light. Therefore, how to comprehensively consider the influence between water, wind and light and jointly predict the power generation power of water, wind and light is a problem that needs to be considered to further improve the accuracy of water power, wind power and photovoltaic power prediction.

[0004] Therefore, a water, wind and light power generation joint prediction method is needed to improve the accuracy of water, wind and light power generation prediction. SUMMARY

[0005] To solve the above problems, the present disclosure provides a water, wind and light power generation joint prediction method and device based on a hybrid neural network.

[0006] In a first aspect, the water, wind and light power generation joint prediction method based on a hybrid neural network comprises:

[0007] Obtaining a multi-source weather observation data set of a target area;

[0008] Obtaining a historical power generation data set of a water, wind and light power station in the target area;

[0009] According to the multi-source weather observation data set and the historical power generation data set, a GRU-CNN water, wind and light power generation joint prediction model is established and trained;

[0010] Real-time acquisition of meteorological forecast data of the target area in the to-be-predicted time period, and transmission to the trained GRU-CNN hydro-wind-solar power generation power joint prediction model to obtain the hydroelectric power, wind power and photovoltaic power of the hydro-wind-solar power station in the target area in the to-be-predicted time period.

[0011] Further, the multi-source meteorological observation data set includes: hydro-wind-solar power station meteorological observation data, satellite meteorological observation data, radar meteorological observation data, ground meteorological observation data, aircraft meteorological observation data and sounding meteorological observation data.

[0012] Further, acquiring the multi-source meteorological observation data set of the target area further includes:

[0013] Numerical prediction assimilation is performed on the multi-source meteorological observation data set of the target area.

[0014] Further, the numerical prediction assimilation of the multi-source meteorological observation data set of the target area includes:

[0015] The initial field of the multi-source meteorological observation data set is assimilated and processed to obtain the processed multi-source meteorological observation data set;

[0016] The WRF regional model is used for prediction to obtain numerical prediction data;

[0017] The numerical prediction data is post-processed based on the processed multi-source meteorological observation data set by using a machine learning algorithm.

[0018] Further, the assimilation and data processing of the initial field of the multi-source meteorological observation data set include:

[0019] The multi-source meteorological observation data set is subjected to the following quality control, and the quality-controlled meteorological observation data set is obtained:

[0020] Repeatability check, validity test, persistence test, extreme value check, position consistency check and spatiotemporal consistency test.

[0021] Further, the repeatability check is used to determine whether there are multiple sets of approximate detection data within the time, latitude, longitude and height limited range, and to remove the repeated data.

[0022] Further, the validity check is used to determine the availability of the current message based on the observation data.

[0023] Further, the persistence test is used to check the observation data caused by instrument failure that does not change or changes less than a threshold value, and when the difference between adjacent message data is less than a given criterion for three or more consecutive times, it is determined that the instrument failure does not pass the persistence test.

[0024] Further, the extreme value check is used to detect whether the observation value of the observation element of the ground area station is within the extreme value range under the corresponding latitude and height conditions.

[0025] Specifically, the threshold range of meteorological observations usually depends on various factors, including geographical location, season, terrain, altitude, etc.

[0026] For example, temperature: the range of temperature is very wide, depending on geographical location and season. For example, in the Arctic region, the temperature in winter may be as low as several tens of degrees below zero; while in the tropical region, the temperature in summer may be as high as 40 degrees Celsius or more.

[0027] Humidity: the range of humidity is usually between 0% (completely dry) and 100% (saturated). However, in actual observation, due to the influence of various factors, humidity rarely reaches these two extreme values.

[0028] Atmospheric pressure: atmospheric pressure is usually expressed in units of hectopascal (hPa). Under standard atmospheric conditions, the atmospheric pressure at sea level is about 1013.25 hPa. With the increase of altitude, the atmospheric pressure gradually decreases.

[0029] Wind speed: wind speed is usually expressed in units of meters per second (m / s) or kilometers per hour (km / h). In calm weather conditions, wind speed may be very low, close to zero; while in extreme weather conditions such as storms or hurricanes, wind speed may be as high as several hundred kilometers per hour.

[0030] Precipitation: the range of precipitation depends on geographical location and season. In arid regions, precipitation may be very little, even no precipitation for consecutive years; while in tropical rainforest regions, precipitation may be very abundant, reaching tens of millimeters or even more per day.

[0031] Further, the position consistency check is used to exclude data quality problems caused by time collection and positioning errors, and to determine whether the current message maintains consistency in time, latitude, longitude, and height changes by comparing the previous and subsequent messages.

[0032] Further, the spatiotemporal consistency check is used to detect whether the message maintains consistency in the process of time and space changes.

[0033] Specifically, the time consistency check:

[0034] Definition: Time consistency refers to the consistency of the state or information of the same entity or resource at different time points or time periods.

[0035] Application: In message transmission, time consistency check can ensure that the timestamp, time expression, etc. of the message remain consistent between different nodes. For example, in a joint test system, the state or information of entity resource A between nodes m and n should remain consistent at the same time T.

[0036] Detection method: Time consistency can be ensured by comparing the timestamps of messages received on different nodes or using a unified time standard (such as UTC time). If a significant difference in timestamp or inconsistent time expression is found, it may indicate that the message has a problem during transmission, affecting its reliability.

[0037] Spatial consistency check:

[0038] Definition: Spatial consistency refers to the consistency of the state or information of the same entity or resource at multiple different spatial locations or nodes.

[0039] Application: In message transmission, spatial consistency check can ensure that the content, structure, etc. of the message remain consistent when transmitted on different network nodes. For example, in a joint test system, the four-dimensional space-time coordinates of entity resource A between nodes m and n should be the same.

[0040] Detection method: Spatial consistency can be ensured by comparing the content, structure, etc. of messages received on different nodes. If changes or damage to the message are found during transmission, it may indicate that there is a problem with the message during spatial transmission, affecting its reliability.

[0041] Further, the numerical prediction data is post-corrected based on the processed multi-source meteorological observation data set by a machine learning algorithm, including:

[0042] Feature extraction, feature preprocessing, feature classification construction, and feature combination are performed on the numerical prediction data.

[0043] The numerical prediction data is post-corrected based on the feature combination of the numerical prediction data and the processed multi-source meteorological observation data set by a machine learning algorithm.

[0044] Further, the GRU-CNN water, wind, and light power generation joint prediction model includes: a heterogeneous information encoding layer, a spatio-temporal feature fusion layer, and a power decoding prediction layer connected in sequence.

[0045] The heterogeneous information encoding layer is used to encode the heterogeneous information in the multi-source meteorological observation data and historical power generation data of the water, wind, and light power station, and generate a spatial matrix of water energy, a spatial matrix of wind energy, and a spatial matrix of light energy.

[0046] The spatio-temporal feature fusion layer is used to generate a spatio-temporal feature matrix of water energy, a spatio-temporal feature matrix of wind energy and a spatio-temporal feature matrix of light energy based on a spatial matrix of water energy, a spatial matrix of wind energy and a spatial matrix of light energy;

[0047] The power decoding prediction layer is used to perform an upsampling operation by inverse convolution based on the spatio-temporal feature matrix of water energy, the spatio-temporal feature matrix of wind energy and the spatio-temporal feature matrix of light energy, then decode the features after inverse convolution based on GRU, and output data through a fully connected layer.

[0048] Specifically, the heterogeneous information encoding layer includes wind and light history information encoding, water and electricity history information encoding and load history information encoding; the spatio-temporal feature fusion layer includes GRU time sequence feature extraction, attention mechanism and CNN spatio-temporal fusion; the power decoding prediction layer is used for GRU decoding and DENSE sequence matching.

[0049] Further, the heterogeneous information includes historical hydroelectric power, historical water level, historical wind power, historical wind speed, historical photovoltaic power, historical irradiance and load data of the water, wind and light power station in the target region.

[0050] Further, the heterogeneous information encoding layer includes a wind power related data encoder, a water power related data encoder, a photovoltaic related data encoder and a load data encoder;

[0051] The input of the wind power related data encoder includes wind power at t-n time, observed wind speed of the water, wind and light power station at t-n time, and forecast meteorological elements of the water, wind and light power station from t time to t-n time.

[0052] The input of the water power related data encoder includes hydroelectric power at t-n time, observed runoff of the water, wind and light power station at t-n time, and forecast meteorological elements of the water, wind and light power station from t time to t-n time.

[0053] The input of the photovoltaic related data encoder includes photovoltaic power at t-n time, observed irradiance of the water, wind and light power station at t-n time, and forecast meteorological elements of the water, wind and light power station from t time to t-n time.

[0054] The input of the load data encoder includes load at t-n time and load at t-n-m time, where t, n and m are natural numbers.

[0055] Further, the generation of the spatial matrix of water energy, the spatial matrix of wind energy and the spatial matrix of light energy includes:

[0056] Based on the heterogeneous information, the spatial features of water energy, wind energy and light energy are extracted by convolution;

[0057] According to the spatial characteristics of water energy, wind energy and light energy, corresponding spatial matrixes of water energy, wind energy and light energy are generated.

[0058] Specifically, the convolutional layer extracts features from the input data through convolution operation. Convolution operation is a mathematical operation that generates new spatial features of water energy, wind energy and light energy by sliding a convolution kernel (also known as a filter) over the input data and calculating the dot product of the convolution kernel and the local region of the input data.

[0059] The rows of the spatial matrix represent the spatial characteristics of the energy corresponding to a position.

[0060] Further, based on the spatial matrixes of water energy, wind energy and light energy, the spatio-temporal feature matrixes of water energy, wind energy and light energy are generated, including:

[0061] Based on the heterogeneous information, the time matrixes of water energy, wind energy and light energy are extracted by GRU;

[0062] Based on the time matrixes of water energy, wind energy and light energy, and the spatial matrixes of water energy, wind energy and light energy, the spatio-temporal feature matrixes of water energy, wind energy and light energy are generated.

[0063] Specifically, the model automatically extracts GRU (Gated Recurrent Unit), a structure of recurrent neural network (RNN), which is particularly suitable for processing sequence data. In terms of feature extraction, GRU can learn and capture long-term dependencies and periodic information in sequence data through its internal mechanism, thereby extracting features useful for specific tasks. That is, the input to GRU is obtained, and the time matrixes of water energy, wind energy and light energy are obtained.

[0064] Further, the spatio-temporal feature fusion layer adopts attention mechanism, sets a weight for each hidden state in the encoder, then weights and sums each hidden state according to the set weight, and finally inputs all the weighted and summed hidden states into the decoder.

[0065] In a second aspect, a hybrid neural network-based water, wind and light power generation power joint prediction device is provided, which includes a data collection unit, a model establishment unit and a prediction unit.

[0066] The data collection unit is configured to obtain a multi-source meteorological observation data set of a target region.

[0067] The data collection unit is also used to acquire historical power generation datasets of hydropower, wind power, and solar power plants in the target area;

[0068] The model building unit is used to build and train a GRU-CNN joint prediction model for hydropower, wind power and solar power generation based on multi-source meteorological observation datasets and historical power generation datasets.

[0069] The prediction unit is used to acquire meteorological forecast data of the target area in real time during the forecast period and transmit it to the trained GRU-CNN joint prediction model for hydropower, wind power and photovoltaic power generation of the target area during the forecast period.

[0070] Thirdly, an electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0071] Memory, which stores computer programs;

[0072] When the processor executes the computer program stored in the memory, it implements the above-mentioned method for joint prediction of hydropower, wind power and solar power generation based on hybrid neural networks.

[0073] Fourthly, a computer-readable storage medium stores a computer program that, when executed by a processor, implements the aforementioned method for joint prediction of hydro, wind, and solar power generation based on a hybrid neural network.

[0074] This disclosure has at least the following beneficial effects:

[0075] This disclosure improves the initial field through data assimilation to enhance numerical forecasting performance. It integrates the spatiotemporal correlation of hydropower, wind power, and solar power generation with load curves to construct an improved GRU-CNN joint prediction model for hydropower, wind power, and solar power generation, thereby achieving joint prediction of hydropower, wind power, and solar power generation and improving the accuracy of hydropower, wind power, and solar power generation prediction.

[0076] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description and the accompanying drawings. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 A prediction method flowchart for the embodiment of the present disclosure is shown in the figure.

[0079] Figure 2 A flowchart for numerical prediction assimilation is shown in the figure.

[0080] Figure 3 A structure diagram for the joint prediction model is shown in the figure.

[0081] Figure 4 A structure diagram for the prediction device of the embodiment of the present disclosure is shown in the figure.

[0082] Figure 5 A structure diagram for the electronic device is shown in the figure. DETAILED DESCRIPTION

[0083] To make the objectives, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be described clearly and completely below with reference to the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are some but not all of the embodiments of the present disclosure. Based on the embodiments in the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present disclosure.

[0084] As shown in the figure, the water, wind and light power joint prediction method based on the hybrid neural network comprises: Figure 1

[0085] S101, obtaining a multi-source meteorological observation data set of a target area.

[0086] S102, obtaining a historical power generation data set of a water, wind and light power station in the target area.

[0087] S103, establishing a GRU-CNN water, wind and light power joint prediction model and training the same according to the multi-source meteorological observation data set and the historical power generation data set.

[0088] S104, obtaining meteorological forecast data of the target area in a to-be-predicted time period in real time and transmitting the same to the trained GRU-CNN water, wind and light power joint prediction model to predict the water power, wind power and photovoltaic power of the water, wind and light power station in the target area in the to-be-predicted time period.

[0089] In specific implementation, the following is introduced:

[0090] S101, obtaining a multi-source meteorological observation data set of a target area.

[0091] ​In some embodiments, the multi-source meteorological observation data set of the target region at least includes water, wind, and light power station observation data, satellite meteorological observation data, radar meteorological observation data, ground meteorological observation data, aircraft meteorological observation data, and sounding meteorological observation data.

[0092] As shown in Figure 2 , the multi-source meteorological observation data set of the target region is numerically predicted and assimilated.

[0093] S102, a historical power generation data set of a water, wind, and light power station in a target region is obtained.

[0094] In some embodiments, the historical power generation data set of the water, wind, and light power station can include historical hydroelectric power, historical wind power, historical photovoltaic power, historical total power, and load.

[0095] S103, an improved GRU-CNN water, wind, and light power generation joint prediction model is established.

[0096] Figure 3 is a structure diagram of the improved GRU-CNN water, wind, and light power generation joint prediction model according to some embodiments of the present specification, as shown in Figure 3 In some embodiments, the improved GRU-CNN water, wind, and light power generation joint prediction model at least includes:

[0097] The heterogeneous information encoding layer is used to encode the heterogeneous information in the multi-source meteorological observation data set and the historical power generation data set of the water, wind, and light power station for numerical prediction and assimilation, and generate a spatial matrix of water energy, a spatial matrix of wind energy, and a spatial matrix of light energy, wherein the heterogeneous information includes historical hydroelectric power, historical water level, historical wind power, historical wind speed, historical photovoltaic power, historical irradiance, and load data of the water, wind, and light power station in the target region, and the heterogeneous information encoding layer can include a wind power related data encoder, a water power related data encoder, a photovoltaic related data encoder, and a load data encoder, wherein the input of the wind power related data encoder can include wind power at t-n time, observed wind speed of the water, wind, and light power station at t-n time, and predicted meteorological elements of the water, wind, and light power station from t time to t-n time, the input of the water power related data encoder can include hydroelectric power at t-n time, observed runoff of the water, wind, and light power station at t-n time, and predicted meteorological elements of the water, wind, and light power station from t time to t-n time, the input of the photovoltaic related data encoder can include photovoltaic power at t-n time, observed irradiance of the water, wind, and light power station at t-n time, and predicted meteorological elements of the water, wind, and light power station from t time to t-n time, and the input of the load data encoder can include load at t-n time and load at t-n-m time.

[0098] The space-time feature fusion layer is configured to generate a space-time feature matrix of water energy, a space-time feature matrix of wind energy and a space-time feature matrix of light energy based on the space matrix of water energy, the space matrix of wind energy and the space matrix of light energy.

[0099] The power decoding prediction layer is configured to perform up-sampling operation by deconvolution, unify data dimensions, decode the features after deconvolution based on GRU, and output data through a fully connected layer, so that the output prediction time meets business requirements, such as ultra-short-term, short-term, medium-term and long-term business requirements.

[0100] The process from the input layer to the hidden layer can be understood as an encoding process, that is, the original data is represented by a set of data with lower dimensions. Since the differences between different forms of energy such as water, wind and light are large, after encoding such heterogeneous information, not only the data can be reduced and the information can be compressed, but also the representative features and the relationships between them can be better extracted, laying a solid foundation for subsequent neural network simulation.

[0101] In some embodiments, the space matrix of water energy, the space matrix of wind energy and the space matrix of light energy are generated by:

[0102] The spatial features of different water, wind and light power stations of each energy are extracted by convolution.

[0103] The space matrix of water energy, the space matrix of wind energy and the space matrix of light energy are generated based on the spatial features of different water, wind and light power stations of each energy.

[0104] Specifically, the wind-related data encoder, the water-related data encoder, the photovoltaic-related data encoder and the load data encoder extract the spatial features of different stations of each energy by CNN (Convolutional Neural Networks, CNN) to form the feature relationship extraction of each station at the spatial level, and obtain the space matrix of water, wind and light energy respectively.

[0105] In some embodiments, the space-time feature fusion layer generates a space-time feature matrix of water energy, a space-time feature matrix of wind energy and a space-time feature matrix of light energy based on the space matrix of water energy, the space matrix of wind energy and the space matrix of light energy.

[0106] The time matrix of water energy, the time matrix of wind energy and the time matrix of light energy are extracted by the GRU layer.

[0107] The space-time feature matrix of water energy, the space-time feature matrix of wind energy and the space-time feature matrix of light energy are generated based on the time matrix of water energy, the time matrix of wind energy, the time matrix of light energy, the space matrix of water energy, the space matrix of wind energy and the space matrix of light energy.

[0108] As Figure 3 shown, the spatio-temporal feature fusion layer introduces an attention mechanism. Because in the time series problem, each time has different importance for subsequent prediction. At the same time, in the joint total power prediction, different stations also have different contributions to the total power. Therefore, an attention mechanism layer is introduced. The attention mechanism layer transmits all the hidden state information in the encoder to the decoder. If there are N hidden states in the encoder, all the hidden state information needs to be transmitted to the decoder. Before transmitting all the information to the decoder, a weight needs to be set for each of the N hidden states. Then each hidden state is weighted and summed according to the set weight, and then all the weighted and summed hidden states are input into the decoder. In this way, the importance of important time and important station information is improved through the weight mechanism, thereby improving the accuracy of joint prediction.

[0109] Through the above CNN+GRU+Attention operation, the spatio-temporal feature fusion of the three kinds of energy is completed. On this basis, considering the influence of power load on the generation side in the new power system, the characteristics of regional power load are further superimposed in the spatio-temporal fusion matrix. And continue to extract the two-dimensional matrix features through the CNN layer. The purpose of this step is to realize the secondary fusion of the generation side spatio-temporal features, and also hope to capture the correlation features between source and load.

[0110] Recurrent Neural Network (RNN) is a commonly used deep learning algorithm in the field of time series prediction at present. However, its network structure leads to the problem of gradient disappearance and gradient explosion in long-term memory and back propagation. Gate Recurrent Unit (GRU) improves this problem through the mechanism of "gate". This network structure is similar to long short-term memory (LSTM), but the difference is that GRU combines the forget gate and input gate in LSTM into a single update gate, and mixes the cell state and hidden state, making the structure more simple and the calculation more efficient. It has good applicability in capturing time series features.

[0111] GRU updates the hidden state h t-1 of the current moment through the reset gate and update gate two multiplication gates to handle the hidden state h t of the previous moment and the external input information x t of the current moment, thereby updating the hidden state h t The calculation rules for model training are as follows:

[0112] z t= σ(W(z)xt + U(z)h t一1 )

[0113] r t = σ(W(r)x t + U(r)h t一1 )

[0114] h't = tanh(Wx t + r t ⊙ Uh t一1 )

[0115]

[0116] where W (z) , W (r) , W, U (z) , U are the response input weight matrix; h't is the current memory content; ⊙ is the Hadamard operation; σ is the Sigmoid activation function,

[0117] The core idea of the attention model (AM) is to introduce attention weights on the input sequence to evaluate the importance of different input information, which helps the model pay more direct attention to the input information with higher correlation with the expected output, and enhances the accuracy of the model.

[0118] Introducing attention mechanism can further improve the extraction ability of GRU network for time sequence features. However, according to the analysis results of space-time correlation characteristics, the output of regional wind, light and hydropower stations not only has correlation characteristics with the historical output of each station itself, but also has strong correlation characteristics with the historical output of other similar stations in the region, and the output of different types of energy has a more significant difference. Ignoring the spatial characteristics of the output between stations and the difference of different types of energy in the data processing and modeling process may limit the further improvement of the prediction accuracy. Therefore, the improved GRU-CNN model shown in the following figure is proposed, which takes into account the difference in data dimensions between input and output modules, and better mines the mapping relationship between input sequence and output sequence.

[0119] S103 further comprises training the improved GRU-CNN hydro-wind-solar power generation power joint prediction model using the multi-source meteorological observation data set and the historical power generation power data set of the numerical prediction assimilation of the hydro-wind-solar power station.

[0120] In specific implementation, the training process is as follows:

[0121] Model training:

[0122] The model is trained using the training set. During the training process, the model learns how to extract features from the input data and make predictions or decisions.

[0123] Training typically involves multiple iterations (epochs), during each of which the model is run through the entire training set once.

[0124] During training, an appropriate optimization algorithm (such as gradient descent, stochastic gradient descent, etc.) is also selected to update the model's parameters in order to minimize the loss function.

[0125] Hyperparameter tuning:

[0126] Hyperparameters are parameters that need to be set before model training, such as learning rate, batch size, number of epochs, etc.

[0127] The validation set is used to adjust the hyperparameters of the model to find the best model configuration [3].

[0128] Model evaluation:

[0129] The test set is used to evaluate the performance of the model. Evaluation metrics can include accuracy, recall, precision, F1 score, etc., depending on the nature of the problem.

[0130] The generalization ability of the model is evaluated by comparing its performance on the test set with the actual performance.

[0131] Model optimization:

[0132] Based on the performance of the model on the validation and test sets, the model can be optimized, such as changing the model structure, adjusting the hyperparameters, using more complex features, etc.

[0133] The loss function is a regular loss function, such as mean square error loss, cross-entropy loss, etc.

[0134] S104, obtain the weather forecast data of the target area in the to-be-predicted time period.

[0135] By improving the GRU-CNN water, wind and light power joint prediction model based on the weather forecast data of the target area in the to-be-predicted time period, the water, wind and light power of the water, wind and light power station in the target area in the to-be-predicted time period is predicted.

[0136] Figure 2 is a flowchart of numerical prediction assimilation of multi-source meteorological observation data set according to some embodiments of the present specification, as shown in Figure 2 As shown in some embodiments, the numerical prediction assimilation of the multi-source meteorological observation data set of the target area includes the following steps:

[0137] S201, assimilate the initial field of the multi-source meteorological observation data set and process the data to obtain the processed multi-source meteorological observation data set.

[0138] The multi-source meteorological observation data set is subjected to repeatability, effectiveness test, continuity test, and spatio-temporal consistency test, to give the quality-controlled observation data and the corresponding quality label, specifically including:

[0139] Repeatability check: mainly to determine whether there are multiple sets of approximate detection data when the time, latitude, longitude and height are close.

[0140] Effectiveness check: the purpose is to determine the current message availability according to the observation data.

[0141] Continuity check: check the observation data that does not change or changes very little caused by instrument failure. The difference between adjacent messages is required to be greater than the given criterion. If the difference between adjacent messages is less than the given criterion for 3 or more consecutive times, it is considered that the instrument failure does not pass the continuity check.

[0142] Extreme value check: the observation value of the observation element of the ground area station should be within a reasonable range under the corresponding latitude and height conditions.

[0143] Position consistency check: mainly to exclude data quality problems caused by time collection and positioning errors. The previous and subsequent messages can be used to determine whether the current message maintains consistency in time, latitude, longitude and height changes.

[0144] Spatio-temporal consistency check: spatio-temporal consistency check can explore the reliability of the message in the process of time and space change.

[0145] Suspicious data check: suspicious data refers to the message data that is not applicable to the above quality control methods. For example, the number of messages of the same flight is only one or two, or there are too many error messages in the same flight, which makes it impossible to normally control the quality. Such data is marked with a suspicious symbol and further processed manually or by other means

[0146] According to the position of the simulated regional station in the model calculated by the model observation operator, the underlying surface information is collected and compared with the actual situation, and the typical differences are marked, specifically including:

[0147] According to the position of the simulated regional station in the model calculated by the model observation operator;

[0148] Collecting underlying surface information and comparing with actual situation;

[0149] Marking the typical differences (e.g., water-land difference).

[0150] For all stations, especially in harsh environments, the detection of observation errors due to equipment exposure to sunlight, rain, and sand erosion, as well as incorrect observations caused by obstacles within the observation field of view, is performed. Obvious non-meteorological observation data are detected and removed. The continuity and consistency of the data are checked, and the observation data are corrected for bias. Specifically, the following steps are included:

[0151] Station extreme value check: Station extreme value refers to the maximum and minimum values of a certain element at the detected station in history.

[0152] Time consistency check: The purpose of the time consistency check is to test the time variation rate of the observation information or observation element, and to identify sudden changes that are not ideal.

[0153] Spatial consistency check: Spatial consistency check is a method of quality control of observation data at the station by fully utilizing the relationship between the observation data at the station and the observation data at multiple adjacent stations at the same time.

[0154] Background field consistency check: The difference between the observation data and the background field (referred to as observation residual error) is compared with the criterion.

[0155] Satellite data needs to be converted into PrepBUFR or BUFR (Binary Universal Form for the Representation of meteorological data) format. In the BUFR format, in addition to the brightness temperature of each observation channel of the satellite, the data grid latitude and longitude, solar elevation angle, solar azimuth angle, satellite azimuth angle, brightness temperature variance, and clear sky ratio are also required. Among them, the brightness temperature variance and the clear sky ratio need to be calculated according to the number of dilution grids. Through the BUFR program attached to the GSI module, FY-2F observation data can be converted into BUFR format.

[0156] The resolution of FY-2F satellite raw data is at least 5 km (infrared channel), which is higher than the resolution of GSI assimilation (0.06°). Through sparse processing, the error correlation between adjacent observations can be reduced. Referring to the BUFR data information of GOES satellite, 9*9 pixel points are taken as an observation unit, and the values of brightness temperature, latitude and longitude, and solar elevation angle are taken from the center point. The brightness temperature variance and clear sky coverage of the observation unit are calculated. Therefore, the resolution of the BUFR data is changed to 40-60 km.

[0157] Under clear sky conditions, only the emission and absorption of the atmosphere and the earth's surface are considered in the infrared band, and scattering effects are not considered. The radiative transfer equation can be written as:

[0158]

[0159] where I is the intensity of radiation received at the top of the atmosphere, I s is the emission from the surface, τ represents the optical thickness in the vertical direction from the surface to the top of the atmosphere, is the optical thickness in the vertical direction for the whole atmosphere, μ is the cosine of the zenith angle, B is the intensity of radiation emitted by the atmosphere. The first term of the radiation transfer equation is the portion of the radiation emitted by the surface that is absorbed by the whole atmosphere and reaches the top of the atmosphere, and the second term is the contribution of the radiation emitted by the atmosphere itself. The portion of the radiation emitted by the atmosphere is obtained by stepwise integration, and the whole equation can be rewritten as:

[0160]

[0161] where, is the change of transmittance with height, which is called the weight function. It can be regarded as the weight of each layer of the atmosphere in the radiation contribution of the whole atmosphere. To some extent, the height of the peak of the weight function corresponds to the height that can be "seen" by the satellite. Most of the satellite observation information is contributed by the height with a larger weight function. The distribution of the weight function is not only related to the waveband, but also related to the gas composition, content and atmospheric state profile in the atmosphere.

[0162] Since the infrared channel radiation information is very sensitive to clouds, cloud detection is an important part of the quality control of infrared channel radiation data. The nominal disc data of FY-2F contains the cloud classification information calculated by the satellite meteorological center DPC system in the non-encrypted observation time. In order to remove the data polluted by clouds as much as possible, while avoiding mixing the model information with the observation field information, three independent cloud detection algorithms that do not depend on the model calculation are used, and cloud detection is carried out by giving empirical thresholds. The detection methods are as follows: when the i grid point satisfies and it is judged as a pixel point with clouds. Where γ represents the vertical decrease rate of atmospheric temperature, and the value is 7 K / 1000 m. is the standard deviation of the terrain height, is the standard deviation of the IR1 brightness temperature (taking 3*3 pixel points centered on i). respectively represent the brightness temperature of the i pixel point of the IR1 and IR4 channels, is the maximum brightness temperature value of the IR1 channel in 3*3 pixel points. The empirical thresholds corresponding to the land and sea surface are different; for the IR1IR4T cloud detection method, since the IR4 near-infrared is easily affected by solar radiation during the day, the empirical threshold ε3 is different during the day and at night. When the solar elevation angle of a certain point is greater than 0, the empirical threshold is taken as ε 31 , and when the solar elevation angle is less than 0, the empirical threshold is taken as ε 32 , the empirical thresholds ε1, ε2, ε 31 , ε 32 are shown in Table 1.

[0163] Table 1

[0164]

[0165] The assimilation of radar is also an important part of data assimilation. The quality of radar data caused by terrain shielding, environmental noise, etc. is checked and controlled. The quality problems such as ground clutter, data shortage caused by other reasons, and clear sky echo are processed by methods such as fuzzy logic method, and high-quality radar data after quality control is given. The data shortage is interpolated and extended: the continuity of meteorological data is processed, and the data obtained at different time periods of the same unit point is interpolated. The abnormal point echo is processed: the echo of Doppler radar has a certain threshold. Under normal circumstances, the echo will not exceed this threshold. The abnormal point data exceeds the echo threshold in many cases. Therefore, this can be used as a basis for determining abnormal points. For the processing of abnormal points, two-dimensional median filtering method is often used. The clear sky echo is removed: sometimes the Doppler radar will also produce echoes when detecting the cloud-free area in the atmosphere. The terrain shielding is processed: most of the echoes caused by terrain shielding are fixed and unchanged. The terrain echo is identified and processed, and the echoes caused by terrain shielding in the radar data are removed. The longitude, latitude and height of the grid point in the Cartesian coordinate system are calculated to obtain the elevation, azimuth and slant range in the spherical coordinate system. Then, according to the calculated elevation, azimuth and slant range in the radar spherical coordinate system, the nearest neighbor method is used for radial interpolation, azimuth interpolation, and the vertical direction is interpolated by linear interpolation method to assign values to the grid point and obtain the analysis value of the grid point. The three-dimensional grid reflectivity data is constructed to analyze the data assimilation system. Based on the numerical prediction results, a rapid cycle analysis is carried out. Based on the Bayesian prior estimation model and the Gaussian probability distribution model, the GSI assimilation module is established to assimilate the collected conventional / irregular observation data. Through the variable scale iteration technique, the reflectivity direct assimilation technique, the wind field weak constraint technique, the cloud analysis false observation and particle information constraint assimilation technique, and the convective scale control variable technique, the analysis field quality is improved. A more optimal initial field is provided for the next forecast.

[0166] The prediction sample of the background error covariance is constructed by the NMC method. The sample is obtained by different starting times.

[0167] The influence of different control variables on assimilation is compared to select the optimal one, and the background error covariance matrix suitable for the climatic scale characteristics of the simulation area is constructed. At the same time, the variable analysis of the finally statistical background error covariance matrix is carried out to obtain its variation at different model heights.

[0168] Variable conversion: a calculation model is established to convert the model prediction variables into control variables.

[0169] Latitude averaging: The difference in the control variable of latitude is averaged by filtering and other methods.

[0170] Inter-variable decorrelation: The correlation between the background errors of variables is removed by dynamic and statistical balance, and converted into independent control variables.

[0171] Vertical transformation: The correlation in the vertical direction of the control variable is removed by empirical orthogonal function (EOF) decomposition, and the eigenvalues and eigenvectors of the background error are calculated.

[0172] Horizontal transformation: The horizontal characteristic length scale of the control variable is calculated by recursive filtering, which represents the characteristics of the background error field in the horizontal direction.

[0173] After strict quality control, the collected conventional / unconventional observation data needs to be converted into a specific format and assimilated into the GSI assimilation system.

[0174] Select the strictly quality-controlled conventional / unconventional observation data of the forecast or simulation period, and pick out the observation data of the 1.5h time domain before and after the assimilation time.

[0175] The selected observation data may have various storage formats such as ASCII, BUFR, MADIS, and programs are written to process all observation data into LITTLE_R format.

[0176] Run the OBSPROC data preprocessing program in the GSI assimilation system to convert the observation data in LITTLE_R format to PREBUFR format for assimilation.

[0177] For different types of observation data, considering different weather backgrounds, the observation operator is established which can reasonably map the relationship between the atmospheric state physical quantity and the observation physical quantity. Taking the establishment of the near-surface observation operator of surface observation data as an example, surface observation data is greatly affected by terrain and topography, and there is a certain height difference between the model terrain and the actual observation station terrain. How to effectively solve the problem of height difference between model terrain and observation station terrain becomes a basic problem of surface data assimilation. Therefore, considering the establishment of a surface observation operator based on the dynamic and thermal constraints of the near-surface boundary layer is an effective way to solve the above problems.

[0178] A new regional near-surface observation operator is established by considering the height difference between the model terrain and the observation station terrain and incorporating the dynamic and thermal processes of the boundary layer.

[0179] The established near-surface observation operator is strictly tested and tested, including the accuracy test of the tangent linear model and the companion model (companion test and gradient test), and the test results ensure that the new near-surface observation operator can better describe the role of near-surface observation in assimilation.

[0180] Different models of terrain and the height difference between the observed terrain result in different forecast effects, so it is necessary to determine the selection of the optimal critical height difference.

[0181] Subsequent assimilation experiments were conducted using the newly established observation operator to ensure that the three-dimensional variational analysis system under the new ground observation operator could accurately reflect the interaction between the wind field and other variables.

[0182] For a certain type of observation data and for a certain forecast target, an integer multiple of the horizontal radius dx of the model grid is selected as the horizontal influence radius for assimilation experiments.

[0183] Increasing or decreasing the horizontal radius of influence allows for assimilation experiments.

[0184] The assimilation was performed several times, and the analytical fields obtained after assimilation under different levels of influence radius were compared with the observation field and background field before assimilation. The assimilation effect of different levels of influence radius was evaluated by incremental or bias analysis. The experiment with the best assimilation effect was selected based on the comparison, and the selection range of the horizontal influence radius was determined.

[0185] Assimilation experiments were conducted by varying the scale of the vertical influence, and the assimilation effects were compared to determine the range of the vertical influence radius.

[0186] Based on the Bayesian prior estimation model and the Gaussian probability distribution model, a 3DVAR assimilation module is established to assimilate the collected conventional / unconventional observation data.

[0187] The initial meteorological field processed by WPS, or the meteorological field predicted by WRF, is used as the background field for three-dimensional variational assimilation.

[0188] A reasonable background error covariance matrix is ​​selected statistically. The preprocessed observation data, background data, and background error covariance are linked with specific names. The GSI assimilation system is run to obtain the assimilated analysis field at the corresponding time, which is used as the background field for the next time.

[0189] Based on the set assimilation time window, multiple cycles of assimilation are performed to continuously improve the forecast results.

[0190] In the assimilation analysis process, for the same type of observation data, for different forecast targets, the scale factor changes in different iteration cycles to fully extract the effective information of the observation data at different scales.

[0191] For a specific type of observational data and a specific forecast target, at the initial analysis time, a value within the previously determined horizontal or vertical radius of influence is selected as the horizontal or vertical scale of influence.

[0192] As weather systems evolve, the scale of influence of observational data at different analysis times may not be the same. The analytical field obtained after the initial assimilation is used as the background field for the next forecast. For the next assimilation time, an assimilation experiment is conducted using the same horizontal or vertical radius of influence as in the previous assimilation. Simultaneously, the radius of influence is increased or decreased in a cyclical manner for assimilation experiments. The assimilation effects under different scales are compared, and the scale of influence corresponding to the set of experiments whose analytical field most closely matches the observations is selected as the horizontal or vertical scale of influence for assimilation at that time.

[0193] For each subsequent analysis time, the optimal horizontal or vertical radius of influence at that time is sought for assimilation, thereby fully extracting information from the observation data at different scales.

[0194] For variable-scale assimilation analysis, multiple iterations are required. Multithreading and parallel computing techniques are employed to improve the computational efficiency of the algorithm and accelerate the analysis of large amounts of data. This enables the system to meet the needs of rapid short-term analysis.

[0195] We consulted literature to find mathematical models for structural optimization design based on system reliability.

[0196] An improvement is made to the optimization criterion method used in the past when solving optimization models.

[0197] The solution process of the examples verifies the good effect of the improved optimization criterion method and ensures faster iterative convergence.

[0198] A multi-level observation operator encompassing various air particle phases is directly employed, avoiding reflectance inversion. This avoids errors introduced during the inversion process and also avoids the problem that warm cloud parameterization schemes, when used as observation operators, cannot accurately reflect ice phase processes. Background temperature is used as the basis for particle classification, enabling the assimilated condensate to more accurately reflect actual observations.

[0199] The updated thermodynamic variables after assimilation of observational data include water vapor mixing ratio, disturbance pressure, disturbance potential temperature, and disturbance geopotential height, but do not include rainwater mixing ratio, which is closely related to the radar's basic reflectivity factor. Therefore, in order to enable the direct assimilation of Doppler weather radar echo data, the rainwater mixing ratio is introduced as an analytical variable in the calculation within the assimilation framework.

[0200] The echo intensity is related to the rainwater mixing ratio based on the relationship between radar reflectivity Z (dBZ) and rainwater mixing ratio qr.

[0201] By introducing a warm rain scheme as a constraint in the GSI assimilation system, the rainwater mixing ratio and the total liquid water mixing ratio are linked.

[0202] Based on the above relationships, direct assimilation of radar echo intensity can be achieved.

[0203] In the GSI objective function, a divergence constraint term is added to the stream function. This term will act as a weak constraint to provide a balance relation constraint to the wind analysis result, and improve the wind analysis result.

[0204] The principle of GSI assimilation is to solve the minimum value of the cost function composed of background field, observation field, analysis field, observation operator, background error covariance, and observation error covariance. Without any modification, the original cost function is used for assimilation test.

[0205] A small term related to the divergence wind is added to the original objective function to perform assimilation test with the balance constraint of the wind field.

[0206] The original scheme and the assimilation results after changing the objective function are compared to ensure that the added constraint term is beneficial to the wind field analysis result. On this basis, the constraint term is continuously improved until the wind field analysis result is optimal.

[0207] A cloud analysis module based on cloud physics initialization is constructed. Cloud physics initialization is to extract information about clouds (mainly cumulus and stratocumulus) from ground truth, sounding, satellite, and radar observations to improve the three-dimensional structure of clouds, and to adjust the atmospheric humidity, temperature distribution, vertical and horizontal wind field, so that the initial state of the model is closer to the real situation of the atmosphere. Through an empirical function, the relative humidity in the initial field is converted into cloud coverage on the corresponding grid to obtain a three-dimensional cloud coverage background field. The cloud base height and cloud coverage information in the mesoscale observation data are added to obtain a comprehensive horizontal cloud field analysis.

[0208] The cloud top is obtained from satellite infrared data, and the radar reflectivity is mainly used to obtain information about clouds in the middle troposphere. Visible light satellite cloud images are used to estimate the total cloud amount and calibrate the overestimation of clouds. The cloud top information is obtained by adding infrared satellite cloud image data, and the three-dimensional cloud coverage obtained from the previous two steps is added to the initial temperature field to realize the correction of cloud analysis, so that the grid analysis cloud top brightness temperature tends to be consistent with the observation. Radar echo data is added. First, the reflectivity factor is interpolated to the model grid points. Within the radar scanning range, the reflectivity factor value at the grid point is compared with the threshold value. If the reflectivity factor is lower than the threshold value, it is considered as clear sky, and if it is higher than the threshold value, the cloud base and cloud top are corrected. Satellite visible light cloud images are added to estimate the total cloud amount and calibrate the overestimation of clouds, and the three-dimensional cloud amount is comprehensively corrected to obtain the final model three-dimensional cloud amount distribution.

[0209] Through cloud analysis, multiple variables can be obtained in the area with cloud distribution, which will be false observations into 3DVAR assimilation, unlike traditional cloud analysis. Cloud analysis information will be a constraint information of model assimilation analysis, into the assimilation system. Through this analysis means, the lack of cloud and rain information in the model and the information of heat, humidity, etc. in the cloud and rain area can be better supplemented. At the same time, in 3DVAR, due to the constraint of model background field and background error covariance matrix, the analysis variables will be more coordinated. Through the cloud analysis module, the strictly corrected model three-dimensional cloud distribution map is obtained. The humidity, cloud water, rain water, temperature and other variables are extracted from the cloud distribution area. These variables are assimilated into the 3DVAR assimilation system as false observations as a constraint information of assimilation analysis, so that the lack of cloud and rain information in the model and the information of heat, humidity, etc. in the cloud and rain area

[0210] The control variables and analysis variables are unified as u, v, w, t, q, and the stream function and potential function are no longer used as control variables. Such analysis results can better show the fine features of mesoscale and small scale. For any forecast object, the environmental field and mesoscale and small scale features at each analysis time are analyzed. For the convective scale weather system, in the GSI assimilation system, the control variables and analysis variables are uniformly set as u, v, w, t, q. The assimilation results are analyzed to see whether the fine features of mesoscale and small scale are better shown.

[0211] From the daily 1200 UTC cold start, every 3 h update three-dimensional variation assimilation, assimilation analysis in the mode of 3 regions at the same time, assimilation including analysis before and after 1.5 h assimilation window multi-source observation data(including routine ground, ship, buoy, automatic station, aircraft message, sounding, etc. Observation data) and satellite, radar and other unconventional observation data, and then 18 h forecast, to the second day 1200 UTC restart stop. The start of the numerical prediction model needs driving field, and the 3-hourly prediction needs to update the driving field in time, and this module is responsible for real-time monitoring of driving data update, collecting driving data for the start of the numerical prediction model. Through shell script, monitor the update of ECMWF global forecast field, get the latest driving data information. Through shell script, start the download software, automatically download the driving data, and store the data in classified and dated backup, which is convenient for future data calling. According to the business needs, determine the spatial range of the mode area and the mode resolution, consider whether to set the nested domain for large-scale fine prediction, determine the appropriate projection method according to the business requirements. Set the mode simulation start and end time. If the 3-hourly start system, update the time setting in the mode through shell script. The start of the mode needs the underlying information as the lower boundary condition, and this module interpolates the global terrain static data available to the simulation area. Read the simulation area information in the mode preprocessing module, calculate the latitude and longitude and the scale factor of each grid point in the map. According to the interpolation method of each variable, interpolate soil type, land pass, terrain height, monthly vegetation cover and monthly albedo, etc. Variables to the mode grid.

[0212] The numerical prediction model needs specific variable data and data format to start, and the output results or reanalysis data of global model as the driving field of mesoscale model, due to its special storage format and various variables, cannot directly drive the numerical model, so it needs to be decoded and extracted through this module. Select the coding table suitable for the driving field. The commonly used NCEP and ECMWF published data adopts GRIB format, and both formats have their own coding table, namely Vtable file. Before driving field extraction, first select the applicable Vtable. Driving variable extraction. Through the determined coding table, determine which fields need to be extracted from the GRIB file, and write the extracted file into the intermediate file of the transition format

[0213] The extracted meteorological elements are horizontally interpolated to the simulation area determined in the model preprocessing module, and written in a data format that can be directly absorbed by the model. The TBL file is read. The module controls how to interpolate meteorological elements through the TBL file. The TBL file provides an interval for each meteorological element, within which the interpolation method of the element field is determined. The interpolated meteorological elements are written. The interpolated meteorological elements are usually written in a data format that can be directly absorbed by the model, usually in NetCDF format, to facilitate visualization by visualization software.

[0214] Global model background field data is processed to the grid of regional numerical prediction through horizontal and vertical interpolation to produce side boundary conditions for regional numerical prediction. The driving field data is read, and the soil field for the model is prepared (usually interpolated to the required height), and the soil classification, land passage, soil temperature, and surface temperature are verified to be consistent with each other. After verifying the soil field, the high-altitude variables are interpolated to the model calculation surface, i.e., the set vertical layers, to generate initial conditions and side boundary conditions.

[0215] The focus is on the occurrence and development of small and medium scale high-impact weather systems. The model settings are the same as above. Through multiple sensitivity tests, the parameterization scheme suitable for the simulation area is obtained. The initial conditions and side boundary conditions generated in the side boundary processing module are read, and the model is started to generate simulation results. The simulation results are sorted and backed up.

[0216] S202, using the WRF regional model to make a prediction to obtain numerical prediction data.

[0217] S203, post-processing the numerical prediction data based on the data-processed multi-source meteorological observation dataset through a machine learning algorithm.

[0218] In some embodiments, S203 can specifically include:

[0219] Feature extraction, feature preprocessing, feature classification construction, and feature combination are performed on the numerical prediction data.

[0220] The numerical prediction data is post-processed based on the feature combination of the numerical prediction data and the data-processed multi-source meteorological observation dataset through a machine learning algorithm.

[0221] Specifically, the roles of GSI and WRF not only improve the temporal and spatial resolution of numerical prediction, but also further improve the prediction accuracy of the model. However, on this basis, a machine learning method is used to make more accurate meteorological element predictions for target stations, in order to lay a good foundation for the following power prediction. The process includes the following steps:

[0222] Download historical observation data, and perform data cleaning on the historical observation data, including effective data filtering, target threshold judgment, date uniqueness judgment, date supplement, and missing data processing;

[0223] Download historical NWP numerical prediction data, and perform feature engineering processing on the NWP data, including feature extraction, feature preprocessing, feature classification construction, and feature combination;

[0224] Input the features and the cleaned observation data into a model for learning, and different models have different learning results. The best model is selected from the learning results, and the model parameters are saved;

[0225] Directly input the NWP data into the previously trained model, and output the corrected single-station numerical prediction product.

[0226] In some embodiments, four machine learning models, namely, Decision Tree (DT), random forest, GBDT (Gradient Boosting Decision Tree), and XGBoost, are used for optimization selection.

[0227] As shown in Figure 4 , the hybrid neural network-based water, wind, and light power generation power joint prediction device comprises a data collection unit 401, a model establishment unit 402, and a prediction unit 403.

[0228] The data collection unit 401 is configured to obtain a multi-source meteorological observation data set of a target region.

[0229] The data collection unit 401 is further configured to obtain a historical power generation data set of a water, wind, and light power station in the target region.

[0230] The model establishment unit 402 is configured to establish and train a GRU-CNN water, wind, and light power generation power joint prediction model according to the multi-source meteorological observation data set and the historical power generation data set.

[0231] The prediction unit 403 is configured to obtain meteorological prediction data of the target region in a to-be-predicted time period in real time, and transmit the meteorological prediction data to the trained GRU-CNN water, wind, and light power generation power joint prediction model to predict water power, wind power, and photovoltaic power of the water, wind, and light power station in the target region in the to-be-predicted time period.

[0232] As shown in Figure 5 , the present disclosure provides an electronic device comprising a processor 501, a communication interface 502, a memory 503, and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 communicate with each other through the communication bus 504.

[0233] The memory 503 stores a computer program.

[0234] The processor 501 is configured to execute the computer program stored in the memory 503 to implement the method described above.

[0235] The present disclosure provides a computer readable storage medium storing a computer program, and the computer program is executed by a processor to implement the method described above.

[0236] The computer readable storage medium can be included in the device / apparatus described in the above embodiments, or can exist separately without being assembled into the device / apparatus. The computer readable storage medium carries one or more programs, and when the one or more programs are executed, the method according to the embodiments of the present disclosure is implemented.

[0237] According to the embodiments of the present disclosure, the computer readable storage medium can be a non-volatile computer readable storage medium, which can include but is not limited to: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any appropriate combination thereof. In the present disclosure, the computer readable storage medium can be any tangible medium that contains or stores a program, which can be used by or in connection with an instruction execution system, apparatus or device.

[0238] Although the present disclosure is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some of the technical features can be replaced by equivalent features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A method for joint prediction of water and wind power generation based on a hybrid neural network, characterized in that, The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The method comprises the following steps: removing the data of the target area polluted by clouds in the weather forecast data of the time period to be predicted, specifically comprising the following steps: 、 and , that is, judging the pixel point as a pixel point with clouds; wherein, represents the vertical decrement rate of atmospheric temperature; is the standard deviation of terrain height, is the standard deviation of IR1 brightness temperature, and 3*3 pixel points are taken with i as the center; 、 respectively represent the brightness temperature of the i-th pixel point in the IR1 and IR4 channels, is the maximum brightness temperature value of the IR1 channel in the 3*3 pixel points. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, wind and light power generation power joint prediction method. The application relates to a hybrid neural network-based water, The numerical prediction data is post-corrected based on the data-processed multi-source meteorological observation dataset through a machine learning algorithm.

5. The hybrid neural network-based joint prediction method of water, wind and solar power generation power according to claim 4, characterized in that, the assimilation and data processing of the initial field of the multi-source meteorological observation dataset comprises: The following quality control is performed on the multi-source meteorological observation dataset, and the quality-controlled meteorological observation dataset is obtained: repeatability check, validity test, continuity test, extreme value check, position consistency check and space-time consistency test.

6. The hybrid neural network-based joint prediction method of water, wind and solar power generation power according to claim 5, characterized in that, the repeatability check is used to determine whether there are multiple sets of detection data within the time, latitude, longitude and height limited range, and to remove the repeated data.

7. The hybrid neural network-based joint prediction method of water, wind and solar power generation power according to claim 5, characterized in that, the validity test is used to determine the availability of the current message according to the observation data.

8. The hybrid neural network-based joint prediction method of water, wind and solar power generation power according to claim 5, characterized in that, the continuity test is used to check the observation data caused by instrument failure, which is unchanged or changes less than a threshold value, and when the difference between adjacent message data is less than a given criterion for three or more times, it is judged that the instrument failure does not pass the continuity test.

9. The hybrid neural network-based joint prediction method of water, wind and solar power generation power according to claim 5, characterized in that, the extreme value check is used to detect whether the observation value of the observation element of the ground area station is within the extreme value range under the corresponding latitude and height conditions.

10. The hybrid neural network-based joint prediction method of water, wind and solar power generation power according to claim 5, characterized in that, the position consistency check is used to exclude data quality problems caused by time collection and positioning errors, and to determine whether the current message maintains consistency in time, latitude, longitude and height changes by comparing the previous and subsequent messages.

11. The hybrid neural network-based joint prediction method of water, wind and solar power generation power according to claim 5, characterized in that, the space-time consistency test is used to detect whether the message maintains consistency in the process of changing with time and space.

12. The hybrid neural network-based joint prediction method of water, wind and solar power generation power according to claim 4, characterized in that, the post-correcting of the numerical prediction data based on the data-processed multi-source meteorological observation dataset through a machine learning algorithm comprises: feature extraction, feature preprocessing, feature classification construction and feature combination of the numerical prediction data; post-correcting of the numerical prediction data based on the feature combination of the numerical prediction data and the data-processed multi-source meteorological observation dataset through a machine learning algorithm.

13. The hybrid neural network-based joint prediction method of water, wind and solar power generation power according to claim 1, characterized in that, the heterogeneous information encoding layer comprises a wind power related data encoder, a water power related data encoder, a photovoltaic related data encoder and a load data encoder; The input of the wind power related data encoder includes wind power at time t-n, observed wind speed of the water-wind-solar power station at time t-n, and forecast meteorological elements of the water-wind-solar power station from time t to time t-n. The input of the water power related data encoder includes water power at time t-n, observed runoff of the water-wind-solar power station at time t-n, and forecast meteorological elements of the water-wind-solar power station from time t to time t-n. The input of the photovoltaic power related data encoder includes photovoltaic power at time t-n, observed irradiance of the water-wind-solar power station at time t-n, and forecast meteorological elements of the water-wind-solar power station from time t to time t-n. The input of the load data encoder includes load at time t-n and load at time t-n-m, where t, n, and m are natural numbers.

14. The hybrid neural network-based water-wind-solar power generation joint prediction method according to claim 1, wherein the generation of the spatial matrix of water energy, the spatial matrix of wind energy, and the spatial matrix of light energy comprises: extracting spatial features of water energy, wind energy, and light energy based on heterogeneous information through convolution; generating the spatial matrix of water energy, the spatial matrix of wind energy, and the spatial matrix of light energy according to the spatial features of water energy, wind energy, and light energy.

15. The hybrid neural network-based water-wind-solar power generation joint prediction method according to claim 1, wherein the generation of the spatiotemporal feature matrix of water energy, the spatiotemporal feature matrix of wind energy, and the spatiotemporal feature matrix of light energy based on the spatial matrix of water energy, the spatial matrix of wind energy, and the spatial matrix of light energy comprises: extracting the time matrix of water energy, the time matrix of wind energy, and the time matrix of light energy based on heterogeneous information through GRU; generating the spatiotemporal feature matrix of water energy, the spatiotemporal feature matrix of wind energy, and the spatiotemporal feature matrix of light energy based on the time matrix of water energy, the time matrix of wind energy, the time matrix of light energy, the spatial matrix of water energy, the spatial matrix of wind energy, and the spatial matrix of light energy. comprises: a data collection unit, a model establishment unit, and a prediction unit; 16. The water and wind power generation power joint prediction device based on a hybrid neural network, characterized in that, the data collection unit is configured to obtain a multi-source meteorological observation data set of a target region; the data collection unit is further configured to obtain a historical power generation data set of a water-wind-solar power station in the target region; the model establishment unit is configured to establish and train a GRU-CNN water-wind-solar power generation joint prediction model based on the multi-source meteorological observation data set and the historical power generation data set; the prediction unit is configured to obtain, in real time, meteorological forecast data of the target region in a to-be-predicted time period, and transmit the meteorological forecast data to the trained GRU-CNN water-wind-solar power generation joint prediction model to predict water power, wind power, and photovoltaic power of the water-wind-solar power station in the target region in the to-be-predicted time period; the GRU-CNN water-wind-solar power generation joint prediction model comprises, in sequence, a heterogeneous information encoding layer, a spatiotemporal feature fusion layer, and a power decoding and prediction layer. ​ ​ The heterogeneous information encoding layer is configured to encode heterogeneous information in multi-source meteorological observation data and historical power generation data of the water, wind and light power station, and generate a spatial matrix of water energy, a spatial matrix of wind energy and a spatial matrix of light energy; The spatio-temporal feature fusion layer is configured to generate a spatio-temporal feature matrix of water energy, a spatio-temporal feature matrix of wind energy and a spatio-temporal feature matrix of light energy based on the spatial matrix of water energy, the spatial matrix of wind energy and the spatial matrix of light energy; The power decoding prediction layer is configured to perform up-sampling operation through deconvolution based on the spatio-temporal feature matrix of water energy, the spatio-temporal feature matrix of wind energy and the spatio-temporal feature matrix of light energy, then decode the features after deconvolution based on GRU, and output data through a fully connected layer; The heterogeneous information includes historical hydroelectric power, historical water level, historical wind power, historical wind speed, historical photovoltaic power, historical irradiance and load data of the water, wind and light power station in the target region; The spatio-temporal feature fusion layer adopts an attention mechanism, sets a weight for all hidden states in the encoder, then weights and sums each hidden state according to the set weight, and inputs all the hidden states after weighting and summing into the decoder; Real-time meteorological forecast data of the target region in the to-be-predicted time period is obtained, including: The method comprises the following steps: removing the data of cloud pollution in the weather forecast data of the target area in the time period to be predicted, specifically comprising the following steps: determining that the pixel point is cloudy when the i-th grid pixel satisfies 、 and ; wherein, represents the vertical decrement rate of atmospheric temperature; is the standard deviation of terrain height, is the standard deviation of IR1 brightness temperature, and 3*3 pixel points are taken with i as the center; 、 respectively represent the brightness temperatures of the i-th pixel point in the IR1 and IR4 channels, is the maximum brightness temperature value of the IR1 channel in the 3*3 pixel points.

17. An electronic device, comprising: The device includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete communication with each other through the communication bus; The memory stores a computer program; The processor is configured to execute the computer program stored on the memory, and realize the method for joint prediction of water, wind and light power generation power based on the hybrid neural network in any one of claims 1-15.

18. A computer readable storage medium storing a computer program, wherein the computer program comprises instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 17. The computer program is executed by the processor to realize the method for joint prediction of water, wind and light power generation power based on the hybrid neural network in any one of claims 1-15.

Images