Wind-light-water cooperative power prediction method capable of adapting to extreme weather change

Abstract

The invention discloses a wind-light-water cooperative power prediction method capable of adapting to extreme weather changes. The method comprises the following core steps: defining a state space containing a plurality of weather modes based on historical hydro meteorological data and power data; dividing historical data into a plurality of data subsets with state tags based on the state space; for each weather state, independently training an exclusive power prediction sub-model by using the corresponding data subset; when real-time power prediction is carried out, firstly, the weather state of the current moment is identified, and the next weather state is pre-judged in combination with the state transition probability matrix; and dynamically calling the sub-models of the current state and the pre-judged next state, and carrying out weighted fusion on the prediction results of the current state and the pre-judged next state to generate a power prediction result. According to the method, through a dynamic switching mechanism of'first classification and second prediction ', extreme weather is effectively dealt with, and the robustness and accuracy of prediction are greatly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of new energy power prediction of power systems, and particularly relates to a wind-solar-water coordinated power prediction method capable of adapting to extreme weather changes. BACKGROUND

[0002] The prior art, such as the short-term wind-solar power prediction method for high-dimensional multi-element meteorological data fusion disclosed in Chinese Patent Application CN116885703A, the core idea of which is to perform comprehensive correlation analysis on power station power generation power and multiple meteorological factor sequences, screen out strong correlation meteorological factors, and construct a power prediction model based on these screened factors. The working principle of this method is based on a basic assumption that there is a stable and universal strong correlation between meteorological factors and power station power generation power.

[0003] However, this method has significant defects when dealing with sudden extreme weather events. Under extreme weather conditions, such as local strong wind, persistent heavy fog, or rainstorm strongly coupled with hydrological processes, the underlying physical relationship between meteorological factors and power generation power will change dramatically and nonlinearly. The strong correlation that originally appeared under normal weather conditions may weaken rapidly, or even reverse. A typical example is the relationship between wind speed and wind power generation power: under normal circumstances, the two are positively correlated, the higher the wind speed, the greater the power; but in extreme gale or gusty weather, the wind turbine will trigger protective shutdown to protect the equipment, resulting in a sharp drop in power generation power to zero, at which time the wind speed and power show strong negative correlation.

[0004] The static and global correlation analysis method used in the prior art cannot capture and distinguish the mode transition of the system dynamic behavior under different weather "states". This leads to poor robustness of the prediction model constructed by the prior art in critical extreme weather periods, and the prediction accuracy will decrease significantly, which cannot meet the demand of high-precision prediction for the safe and stable operation of power grids. SUMMARY

[0005] The purpose of the present application is to provide a wind-solar-water coordinated power prediction method and device capable of adapting to extreme weather changes, to solve the problems in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] A wind-solar-water coordinated power prediction method capable of adapting to extreme weather changes, comprising the following steps: defining a state space containing multiple weather modes based on historical hydro-meteorological data and power generation data; dividing the historical data into multiple data subsets with state labels based on the state space; for each state label, independently training a power prediction sub-model using its corresponding data subset; when performing real-time power prediction, identifying the weather state to which the current time belongs, and calling the power prediction sub-model corresponding to the weather state to generate a power prediction result.

[0008] Preferably, the historical hydro-meteorological data includes reservoir inflow data or basin area rainfall data reflecting large-scale hydro-meteorological processes, and wind speed or illumination intensity rate of change or local variance reflecting local meteorological characteristics.

[0009] Preferably, the step of dividing the historical data into multiple data subsets is implemented using a recursive segmentation algorithm.

[0010] Further, a penalty term is introduced into the optimization objective function of the recursive segmentation algorithm, which is negatively related to the proportion of minority class samples in the split child nodes, so as to retain nodes containing extreme weather state samples.

[0011] Further, before the step of training the power prediction sub-model, it further comprises the step of: for each data subset with a state label, independently performing comprehensive correlation analysis to generate a dedicated strong correlation meteorological factor set for each weather state; the power prediction sub-model is trained based on its corresponding dedicated strong correlation meteorological factor set.

[0012] Further, before the step of training the power prediction sub-model, it further comprises the step of: for the extreme weather state data subset with a sample size less than a preset threshold, using a synthetic data generation technique for time series to expand the data subset.

[0013] Preferably, in the step of dividing the historical data into multiple data subsets, the feature vector used to identify the weather state is constructed by: performing continuous wavelet transform on the time series of key variables in the historical hydro-meteorological data to generate a time-frequency energy matrix; dividing the time-frequency energy matrix into multiple preset frequency bands on the frequency axis; calculating the energy features in each preset frequency band; and splicing all the energy features calculated by all key variables in all frequency bands to form the feature vector.

[0014] Further, the method further comprises: constructing a state transition probability matrix based on the state label sequence of the data subset; when performing real-time power prediction, combining the state transition probability matrix to pre-judge the next weather state in one or more time steps in the future; and weighting and fusing the prediction value of the power prediction sub-model of the current state and the prediction value of the power prediction sub-model of the pre-judged next weather state to obtain the final power prediction result.

[0015] More specifically, the weighting fusion calculates the final power prediction value P f (t) by the following formula:

[0016] P f (t) = w c ·M c (X t ) + w n ·M n (X t )

[0017] wherein w n = P(R n |R c )·e -k / τ and w c = 1-w n ; in the formula, M c is the power prediction sub-model of the current state R c , M n is the power prediction sub-model of the pre-judged next weather state R n , X t is the current input, P(R n |R c ) is the transition probability from state R c to R n , k is the prediction time step, and tau is the time decay constant.

[0018] Further, in the process of identifying the weather state to which the current time belongs, it further comprises: calculating the distance between the feature vector of the current time and each state in the state space to obtain an anomaly score; when the anomaly score exceeds a preset threshold, it is determined that an unknown weather event is encountered, and a preset general robust model is called to generate a power prediction result.

[0019] More specifically, when the anomaly score exceeds the preset threshold, before the general robust model is called, it further comprises: triggering a feature enhancement subprogram to calculate higher-order statistical features or time-frequency domain features of the feature vector and supplement them to the original feature vector for state recognition again.

[0020] Compared with the prior art, the present application has the following beneficial effects:

[0021] 1. This invention introduces state-space partitioning to decompose the complex global prediction problem into multiple local prediction problems corresponding to specific weather states, and trains a dedicated "expert model" for each state. This dynamic switching mechanism of "classification first, prediction later" can adaptively change its internal prediction logic and dependent feature combinations according to real-time weather conditions, effectively addressing the problem of abrupt changes in the physical relationship between power generation and meteorological factors caused by extreme weather, avoiding the failure of a single global model in non-stationary environments, thereby significantly improving the robustness and accuracy of prediction.

[0022] 2. This invention defines the state space by introducing variables that reflect large-scale, long-period hydrometeorological processes, such as inflow from hydropower stations or watershed rainfall. This solves the problem of accurately judging the impact of large-scale weather systems based solely on local wind and solar meteorological data, making state division more accurate and physically meaningful. Within each precisely defined state, data consistency is higher, enabling state-specific correlation analysis and model training to more deeply uncover the physical laws under that specific scenario.

[0023] 3. By constructing a state transition probability matrix and introducing a forward-looking fusion prediction mechanism, this invention can anticipate and respond to impending weather state changes. It not only identifies the current state but also predicts the next possible state, dynamically weighting and fusing the predictions from both. This design, transforming passive "post-event identification" into proactive "pre-event prediction," significantly shortens the model's response time to rapidly changing weather and improves prediction accuracy during dynamic processes.

[0024] 4. By combining an anomaly detection module and an adaptive feature enhancement strategy, this invention establishes a set of...

[0025] The system employs a "recognition-alarm-enhancement-rollback" defense mechanism. When faced with complex extreme weather conditions never seen in the training data, the system can identify anomalies and attempt reclassification by enhancing features. If this fails, it invokes a general, robust model with strong generalization capabilities. This avoids catastrophic prediction biases caused by misjudgments of the state, greatly enhancing the safety and reliability of the entire prediction system in complex real-world environments.

[0026] 5. This invention effectively solves the problem of imbalanced training caused by scarce extreme weather samples by employing an adaptive recursive segmentation algorithm based on sample balancing and state-specific data augmentation techniques. The improved segmentation algorithm protects minority class samples from over-segmentation, while the data augmentation technique expands the training set of rare states. This enables higher performance levels for modeling both normal and extreme states, thereby comprehensively improving overall prediction accuracy. Attached Figure Description

[0027] To more clearly illustrate the technical solution of the present invention, the accompanying drawings are briefly described below:

[0028] Figure 1 This is a schematic diagram of the overall process of a wind-solar-hydro power prediction method that can adapt to extreme weather changes, provided by an embodiment of the present invention.

[0029] Figure 2 This is a detailed flowchart illustrating the state recognition and data segmentation steps in an embodiment of the present invention;

[0030] Figure 3 This is a detailed flowchart illustrating the dynamic fusion prediction step in an embodiment of the present invention;

[0031] Figure 4 This is a schematic diagram illustrating the principle of feature extraction based on continuous wavelet transform in an embodiment of the present invention;

[0032] Figure 5 A schematic diagram of the functional module structure of a water-wind-solar collaborative prediction device based on dynamic state division provided in an embodiment of the present invention;

[0033] Figure 6 for Figure 5 A detailed functional module structure diagram of the dynamic prediction module. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the specific embodiments of this invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the scope of protection of this invention. In drafting the embodiments, it should be ensured that the language is clear, complete, and accurate, and that it supports each technical feature of the claims, so that those skilled in the art can understand and implement this invention.

[0035] Example 1

[0036] This embodiment aims to disclose a wind-solar-hydro power prediction method that can adapt to extreme weather changes, in order to solve the problem of low prediction accuracy of existing technologies under extreme weather conditions.

[0037] Reference Figure 1 The method 100 provided in this embodiment may specifically include the following steps:

[0038] Step S110: Define the state space.

[0039] Step S120: State recognition and data segmentation.

[0040] Step S130: Stateful model training.

[0041] Step S140: Dynamic fusion prediction.

[0042] Step S110: Define state space

[0043] In this step, in order to accurately depict different weather patterns, especially those related to water power generation, a comprehensive state space set R needs to be constructed. Unlike existing technologies that rely only on local meteorological factors, the present application innovatively introduces third-party data sources that can reflect large-scale, long-period hydro-meteorological processes.

[0044] Specifically, the collected historical data not only includes local meteorological data (such as wind speed, light intensity, temperature, humidity, etc.) and power generation data of wind power stations and photovoltaic power stations, but also includes inflow data of hydropower stations or areal rainfall data of the river basin where they are located, which are strongly related to extreme weather (especially precipitation). These large-scale hydrological data can more accurately reflect regional and persistent weather processes, such as persistent heavy rain, thereby making up for the limitations of local meteorological data in spatial and temporal scales.

[0045] Preferably, in order to capture dynamic changes, in addition to the original data, the rate of change and local variance of key variables and other statistical characteristics are also calculated. For example, the rate of change and variance of wind speed and light intensity within a small time window in the past are calculated, and these features can effectively distinguish between smooth weather and episodic, volatile weather.

[0046] Based on the above multi-source heterogeneous data, a state space set R containing multiple typical weather patterns is defined through expert knowledge or clustering analysis. For example, the state space can be defined as R = {normal state, strong wind / weak light state, heavy precipitation / low wind speed state, water wind light collaborative fluctuation state}. Each state corresponds to a specific weather pattern that has a significant impact on water wind light power generation.

[0047] Step S120: State recognition and data segmentation. Refer to Figure 2 This step aims to segment the complete historical time series data set into several data subsets with unique state labels according to the state space defined in step S110.

[0048] 1. Feature vector construction

[0049] In order to accurately identify the weather state to which each time point belongs, a feature vector with rich information needs to be constructed for each time point. In one preferred embodiment of the present application, multi-scale time-frequency energy features based on continuous wavelet transform (CWT) are used to replace traditional macroscopic statistical features. Traditional methods such as calculating mean and variance compress the dynamic process within a time window into static numerical values, losing key time sequence structure information.

[0050] Refer toFigure 4 The feature construction method of the present application is as follows:

[0051] First, for each key meteorological / hydrological variable (such as wind speed, light intensity, inflow), a continuous wavelet transform (CWT) is performed on its time series S(t) at the past T time points to generate a two-dimensional time-frequency energy matrix, also known as a wavelet spectrogram 401. This transform can capture the local features of the signal in both time and frequency dimensions.

[0052] Second, according to meteorological physical knowledge, the wavelet spectrogram 401 is divided into several frequency bands with clear physical meaning on the frequency axis (scale axis). For example, it can be divided into a high frequency band 402 (corresponding to turbulence, gusts, etc. fast fluctuations at the second to minute level), a medium frequency band 403 (corresponding to weather system evolution at the minute to hour level), and a low frequency band 404 (corresponding to slow background changes of daily variation or weather process).

[0053] Third, within each pre-defined frequency band, the energy features such as the total energy or the statistical moments of the energy distribution (such as the energy mean, energy variance) are calculated.

[0054] Finally, all the energy features calculated by all key variables in all frequency bands are spliced together to form a new feature vector that can describe multi-scale dynamic processes. This feature vector retains the dynamic "fingerprint" of the weather process, greatly enhancing the recognition of rare composite weather events with complex internal structure.

[0055] 2. Adaptive recursive segmentation

[0056] After obtaining the feature vector, a recursive segmentation algorithm is used to traverse the entire historical data set, classifying each time point into the most likely weather state. To solve the problem of data imbalance caused by the fact that extreme weather samples are much less than normal weather samples in the historical data, the recursive segmentation algorithm is improved in this embodiment.

[0057] Specifically, at each iteration split of the algorithm, the optimization objective function not only considers the information gain or purity improvement, but also introduces a penalty term. The penalty term is negatively related to the proportion of minority class samples (i.e. extreme weather state samples) in the child nodes after splitting. This means that if a split would result in the nodes containing rare samples being further subdivided, which could cause these rare samples to be ignored or overfit in subsequent training, then this split would be "penalized" and the algorithm would tend to choose other splitting methods. This mechanism causes the algorithm to actively retain nodes containing rare samples when segmenting, avoiding their premature and excessive division, thereby alleviating the overfitting risk of minority class samples and ensuring the quality of the data subsets of each state after segmentation.

[0058] Step S130: State-specific model training. After obtaining the state-labeled data subsets, this step trains an "expert" power prediction sub-model M for each state independently R .

[0059] 1. State-specific data augmentation

[0060] Before training, the sample size of each state data subset is first checked. For the extreme weather state data subsets with insufficient sample size (e.g., the number of samples is less than a pre-set threshold), data augmentation techniques for time series are employed. For example, synchronous, physically-constrained perturbations or interpolations can be performed on multi-dimensional input sequences such as meteorological and hydrological data to generate new, high-quality training samples. This effectively expands the data set for rare states, providing a data basis for training robust expert models.

[0061] 2. State-specific correlation analysis

[0062] Next, for each (possibly augmented) data subset, a comprehensive correlation analysis step is independently performed. This operation aims to generate a list of meteorological factors that are most relevant to power generation in each weather state R, denoted as SD R . For example, in the "strong wind" state, wind speed may be the most critical factor; while in the "heavy precipitation" state, inflow and light intensity may become more important. This state-specific feature selection ensures that each sub-model uses the optimal input variables.

[0063] 3. State-specific model training

[0064] Finally, for each weather state in the state space R, a power prediction sub-model M R is independently trained using its specific set of strongly correlated factors SD R and the corresponding state data subset. These sub-models can be any advanced machine learning models such as gradient boosting decision trees (GBDT), long short-term memory networks (LSTM), etc. Since each model is trained for a specific, relatively stable weather pattern, its learning objective is more explicit and data consistency is higher, thus achieving higher prediction accuracy than a single global model.

[0065] Step S140: Dynamic fusion prediction. Referring to Figure 3 , this step describes the complete process when performing real-time power prediction.

[0066] 1. Anomaly detection and adaptive augmentation

[0067] In real-time prediction, the real-time meteorological and hydrological data at time t is first obtained, and the feature vector at time t is constructed according to the method described in step S120.

[0068] In order to cope with the compound extreme weather that has never occurred in history, the system introduces an anomaly detection module in parallel. The module calculates the Mahalanobis distance or other distance metrics of the current feature vector from the center (or distribution) of all defined states R, and obtains an anomaly score.

[0069] When the score exceeds a preset threshold, the system determines that the current may encounter unknown compound extreme weather. At this time, the system first automatically triggers a feature enhancement subroutine to calculate higher-order statistical features (such as skewness, kurtosis) or other time-frequency domain features, and supplements them to the original feature vector, and then performs state recognition again.

[0070] If the enhanced state still cannot be accurately classified (i.e., the anomaly score is still high), the system will give up calling any expert model M R , and instead call a specially trained and more generalizable “general robust model”. The general model may use a wider set of features and be trained with robustness as the primary goal. This mechanism guarantees the lower limit of prediction in unknown situations and avoids large prediction bias caused by misjudgment of the state.

[0071] 2. State transition prediction and forward-looking fusion

[0072] If the current time is not determined to be abnormal, the system identifies the weather state R c = R(t) to which the current belongs.

[0073] To solve the natural lag problem of state recognition, the embodiment introduces a state transition prediction mechanism. Based on the state label sequence generated after dividing the historical data, a state transition probability matrix can be constructed in advance, which stores the historical conditional probability P(R j |R i ) of transition from any state to another state.

[0074] In prediction, the system not only identifies the current state R c , but also predicts the most likely next state R n in the future one or more steps k by combining the current feature vector and the state transition probability matrix.

[0075] The final power prediction result is obtained by weighted fusion of the prediction value of the expert model M c of the current state and the prediction value of the expert model M n of the predicted next state. The specific fusion calculation is as follows:

[0076] Pf (t) = w c · M c (X t ) + w n · M n (X t ) (1)

[0077] where the weight w c and w n are calculated as:

[0078] w n = P(R n |R c ) · e -k / τ (2)

[0079] w c = 1 - w n (3)

[0080] In the above formula:

[0081] · P f (t) is the final fused prediction power at time t.

[0082] · M c (X t ) and M n (X t ) are the predicted outputs of the current state expert model and the next state expert model based on the current input X t .

[0083] · w c is the weight of the current state model M c , w n is the weight of the predicted next state model M n .

[0084] · P(R n |R c ) is the historical conditional probability of transitioning from the current state R c to the next state R n , which is obtained from the pre-constructed state transition matrix.

[0085] · k is the time step of prediction, for example, predicting 15 minutes in the future, k can be 1.

[0086] · τ is an adjustable time decay constant hyperparameter, used to control the speed of the weight of the prediction model decaying with the increase of the prediction time step k. The larger the τ value, the higher the trust degree of the long-term prediction, and the greater the influence.

[0087] The weighted fusion mechanism introduces an exponential decay factor e -k / τThis achieves a quantitative penalty for the uncertainty of future predictions; that is, the longer the prediction time step k, the higher the predicted next state R. n The greater the uncertainty, the more important it is for the corresponding model M. n weight w n The lower the value, the better. Simultaneously, the fusion weights are directly related to the physically defined state transition probability P(R). n |R c The linkage between weights and weights provides a clear physical basis and explanation, ensuring the logical rigor and operability of the fusion mechanism.

[0088] Example 2

[0089] This embodiment provides a water-wind-solar collaborative prediction device 200 based on dynamic state division, the structure of which is shown in the attached figure. Figure 5 As shown. This device can be a dedicated hardware device or a software system deployed on a server or industrial control computer. The device 200 includes:

[0090] The state space definition module 210 is used to execute step S110 described in Embodiment 1. Specifically, this module is responsible for collecting historical hydrological and meteorological data (including inflow of hydropower stations, watershed rainfall, etc.) and power generation data, and defining a state space containing multiple weather patterns based on these data.

[0091] The data segmentation module 220 is used to perform step S120 as described in Embodiment 1. This module receives historical data and a state space definition, and segments the historical data into multiple data subsets with state labels. Preferably, this module integrates a feature extraction unit 221 (not included in...). Figure 5 As shown in the diagram (but the functionality is included), this unit is configured to perform continuous wavelet transform on the time series of key variables and extract multi-scale time-frequency energy features to construct feature vectors for state identification. Furthermore, this module employs an improved adaptive recursive segmentation algorithm to effectively handle imbalanced data.

[0092] The model training module 230 is used to perform step S130 as described in Embodiment 1. This module receives the segmented data subset and, for each state label, independently performs data augmentation (if necessary), feature selection, and finally trains a dedicated power prediction sub-model.

[0093] The dynamic prediction module 240 is used to execute step S140 as described in Embodiment 1. This module is the core of the system's real-time operation. (Refer to...) Figure 6 The dynamic prediction module 240 may further include:

[0094] • a state recognition unit 241 for receiving real-time input data, constructing feature vectors, and identifying the weather state to which the current time belongs.

[0095] • a state transition prediction unit 242 for predicting the most likely future weather state according to a pre-constructed state transition probability matrix.

[0096] • a model calling and fusion unit 243 for dynamically calling the corresponding current state expert model and next state expert model according to the results of the state recognition unit 241 and the state transition prediction unit 242, and performing weighted fusion according to formula (1) in Embodiment One to generate the final power prediction result.

[0097] Further, the apparatus 200 can further include an anomaly detection module 250. The module works in cooperation with the dynamic prediction module 240. Before state recognition, the anomaly detection module 250 first calculates the anomaly score of the current input. If the score exceeds the threshold, it will notify the dynamic prediction module 240, which will call a pre-set general robust model to generate a prediction, thereby ensuring the safety of the system in unknown situations.

[0098] Embodiment Three

[0099] The present embodiment provides a computer-readable storage medium. The computer-readable storage medium can be any type of volatile or non-volatile storage device, or a combination thereof, such as random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory, hard disk drive (HDD), solid-state drive (SSD), optical disc (such as CD-ROM, DVD), etc.

[0100] In the present embodiment, the computer-readable storage medium has computer program instructions stored thereon. When these computer program instructions are executed by one or more processors (e.g., general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, field-programmable gate arrays FPGA, etc.), the processors can be caused to implement all or part of the steps of the wind-solar-water coordinated power prediction method capable of adapting to extreme weather changes as described in Embodiment One above.

[0101] Specifically, the computer program instructions can include code segments for performing the following operations:

[0102] • an operation of defining a state space containing multiple weather patterns;

[0103] • an operation of dividing historical data into multiple data subsets with state labels based on the state space, which can include feature extraction based on continuous wavelet transform and adaptive recursive segmentation;

[0104] • For each state label, the operation of training power prediction sub-model independently, which may include data augmentation and state-specific feature selection;

[0105] • When performing real-time power prediction, the operation of performing anomaly detection, state identification, state transition prediction, and weighted fusion prediction.

[0106] By storing the computer program implementing the method of the present application on a computer readable storage medium, the method can be conveniently distributed, installed and executed on various computing devices, thereby realizing the technical effects described in the present application.

[0107] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application should not be limited to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be considered as falling within the protection scope of the present application.

Claims

1. A method for predicting the combined power of wind, solar, and hydropower that can adapt to extreme weather changes, characterized in that, Includes the following steps: Based on historical hydrological and meteorological data and power generation data, a state space containing multiple weather patterns is defined. The historical data is segmented into multiple data subsets with state labels based on the state space; For each state label, a power prediction sub-model is trained independently using its corresponding subset of data; When performing real-time power prediction, the current weather state is identified, and the power prediction sub-model corresponding to the weather state is called to generate the power prediction result.

2. The method according to claim 1, characterized in that, The historical hydrological and meteorological data include hydropower station inflow data or basin-wide rainfall data reflecting large-scale hydrological and meteorological processes, as well as the rate of change or local variance of wind speed or light intensity reflecting local meteorological characteristics.

3. The method according to claim 1, characterized in that, The step of dividing historical data into multiple data subsets is specifically implemented using a recursive segmentation algorithm.

4. The method according to claim 3, characterized in that, The optimization objective function of the recursive segmentation algorithm introduces a penalty term, which is negatively correlated with the proportion of minority class samples in the child nodes after splitting, in order to retain nodes containing extreme weather state samples.

5. The method according to claim 1, characterized in that, Before the step of training the power prediction sub-model, the following step is also included: For each data subset with a status label, an independent comprehensive correlation analysis is performed to generate a unique set of strongly correlated meteorological factors for each weather state. The power prediction sub-model is trained based on its corresponding set of strongly correlated meteorological factors.

6. The method according to claim 1, characterized in that, Before the step of training the power prediction sub-model, the following step is also included: For extreme weather data subsets with sample sizes less than a preset threshold, synthetic data generation techniques for time series data are used to expand the data subset.

7. The method according to claim 1, characterized in that, In the step of segmenting historical data into multiple data subsets, the feature vector used to identify weather conditions is constructed in the following way: performing continuous wavelet transform on the time series of key variables in the historical hydrological and meteorological data to generate a time-frequency energy matrix; The time-frequency energy matrix is ​​divided into multiple preset frequency bands on the frequency axis; Calculate the energy characteristics within each of the preset frequency bands; The energy characteristics of all key variables calculated across all frequency bands are concatenated to form the feature vector.

8. The method according to claim 1, characterized in that, Also includes: Based on the state label sequence of the data subset, a state transition probability matrix is ​​constructed; When performing real-time power prediction, the state transition probability matrix is ​​used to predict the next weather state within one or more future time steps. The power prediction sub-model for the current state is weighted and fused with the power prediction sub-model for the predicted next weather state to obtain the final power prediction result.

9. The method according to claim 8, characterized in that, The weighted fusion is used to calculate the final power prediction value P using the following formula. f (t): P f (t)=w c ·M c (X t )+w n ·M n (X t ) Among them, w n =P(R) n |R c )·e -k / τ and w c =1-w n In the formula, M c The current state R c The power prediction sub-model, M n For the predicted next weather state R n The power prediction sub-model, X t Given the current input, P(R) n |R c ) is from state R c Transfer to R n The transition probability is given by k, where k is the prediction time step and τ is the time decay constant.

10. The method according to claim 1, characterized in that, The process of identifying the current weather condition also includes: Calculate the distance between the feature vector at the current moment and each state in the state space to obtain an anomaly score; When the abnormal score exceeds a preset threshold, it is determined that an unknown weather event has occurred, and a preset general robust model is invoked to generate a power prediction result; When the abnormal score exceeds a preset threshold, before calling the general robust model, the process further includes: triggering a feature enhancement subroutine to calculate higher-order statistical features or time-frequency domain features of the feature vector and add them to the original feature vector, and then performing state recognition again.

Images