A new energy power prediction method and system based on a multi-scale state decomposition mechanism

Abstract

A new energy power prediction method and system based on a multi-scale state decomposition mechanism, wherein the new energy power prediction method comprises: collecting multi-source historical data of a wind power station or a photovoltaic power station, preprocessing the multi-source historical data, and then constructing a training data set; constructing a new energy power prediction network, the new energy power prediction network comprising a knowledge guiding layer, a historical sequence encoder, a continuous dynamic system modeling layer, and a physical perception decoding network; the knowledge guiding layer and the historical sequence encoder are connected in parallel, and the outputs of the knowledge guiding layer and the historical sequence encoder are sequentially connected with the continuous dynamic system modeling layer and the physical perception decoding network; the training data set is used to train the new energy power prediction network; the trained new energy power prediction network is deployed to a device end, and power prediction is performed to obtain a prediction result. The present application significantly improves the stability, convergence speed and robustness of the prediction model under extreme weather conditions in the early training stage.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy power prediction, and particularly relates to a new energy power prediction method and system based on a multi-scale state decomposition mechanism. BACKGROUND

[0002] At present, in the complex spatio-temporal sequence prediction task, especially for new energy power prediction, new type of power system source and load prediction, multi-modal weather field modeling and other application fields, there are various mainstream modeling methods, mainly including:

[0003] (1) Statistical and traditional time series modeling methods:

[0004] The traditional time series method mainly relies on the statistical regularity of historical data for modeling, and the typical models include autoregressive moving average model (ARIMA), autoregressive conditional heteroscedasticity model (ARCH / GARCH), Kalman filter, Markov process modeling, etc. These methods can effectively capture linear correlation and short-term trend in short time series, and have certain adaptability for stable and stationary data sequences. However, its modeling ability is mainly concentrated in a single time dimension, and it shows obvious limitations in the face of complex spatial dependence, non-stationary strong nonlinear system, and is difficult to adapt to the cross-scale, multi-modal, long-period coupling dynamic evolution characteristics of new energy power system.

[0005] (2) Traditional deep learning model:

[0006] With the development of deep learning, recurrent neural network (RNN), long short-term memory network (LSTM), gated recurrent unit (GRU), convolutional neural network (CNN) and other models are widely used in time series prediction tasks. Especially in the field of new energy power prediction, through the learning of historical time series and part of the weather data, the model can automatically extract complex nonlinear mapping relationship and improve the prediction accuracy. However, this kind of model generally has the problems of limited spatial feature utilization; lack of expression of spatial continuity and topological structure; limited long-term dependence learning ability, easy to appear gradient disappearance; weak generalization ability for complex scenes, etc.

[0007] (3) Spatio-temporal joint deep learning model:

[0008] To solve the problem of traditional deep model space-time decoupling, in recent years, deep models combining space and time features have appeared, such as ConvLSTM, ST-LSTM, Causal LSTM and the like. These models preliminarily realize the coupling learning of space-time features by introducing spatial convolution operation or bidirectional memory mechanism in the hidden state, and partially improve the expression ability of the model to the spatial dynamic process. Subsequently, multi-modal spatio-temporal models integrating satellite remote sensing, geographic information, and meteorological simulation data began to be applied to new energy system prediction tasks. However, existing methods still mainly stack space and time information at the feature fusion level, and essentially still focus on discrete sequence modeling, lacking the ability to model the intrinsic continuous dynamic rules of the system.

[0009] (4) Application exploration of pre-training large models:

[0010] Recently, with the development of meteorological large models (such as GraphCast, FourCastNet), multi-modal pre-training models (such as remote sensing pre-training models), and domain knowledge graphs, some studies attempt to use the meteorological field prior or scene knowledge provided by large models to embed into the prediction model. However, at this stage, related research is mostly at the level of directly using the output of large models as input features, and there is still a lack of deep integration mechanism between large model knowledge and downstream complex system dynamic modeling. SUMMARY

[0011] The present application provides a new energy power prediction method and system based on a multi-scale state decomposition mechanism to solve the technical problems mentioned in the background art.

[0012] To achieve the above purpose, the technical scheme of the present application is as follows:

[0013] The present application provides a new energy power prediction method based on a multi-scale state decomposition mechanism, comprising the following steps:

[0014] S1, collecting multi-source historical data of a wind power station or a photovoltaic power station, preprocessing the multi-source historical data, and then constructing a training data set;

[0015] S2, constructing a new energy power prediction network, the new energy power prediction network comprising a knowledge guiding layer, a historical sequence encoder, a continuous dynamic system modeling layer, and a physical perception decoding network; the knowledge guiding layer and the historical sequence encoder are connected in parallel, and the outputs of the knowledge guiding layer and the historical sequence encoder are connected to the continuous dynamic system modeling layer and the physical perception decoding network in turn;

[0016] S3, training the new energy power prediction network using the training data set to obtain a trained new energy power prediction network;

[0017] S4, deploying the trained new energy power prediction network to the device end and performing power prediction to obtain a prediction result.

[0018] Further, the S1 specifically comprises the following steps:

[0019] S11, collecting multi-source historical data, the multi-source historical data comprising a historical power output sequence and synchronized meteorological data, if it is a wind power station, the meteorological data comprising environmental factors affecting the output of a wind turbine; if it is a photovoltaic station, the meteorological data comprising factors affecting the power of a photovoltaic component;

[0020] S12, pre-processing the multi-source historical data, the pre-processing at least comprising one or more of time alignment, missing value completion and outlier removal, and then performing normalization processing to obtain pre-processed multi-source historical data;

[0021] S13, constructing a training data set according to the pre-processed multi-source historical data.

[0022] Further, the historical sequence encoder comprises a multi-layer stacked Transformer self-attention encoder, each layer of the Transformer self-attention encoder comprising a multi-head attention mechanism and a feedforward network, the multi-head attention mechanism and the feedforward network being connected through a residual connection.

[0023] Further, the S3 specifically comprises the following steps:

[0024] S31, selecting a feature sequence set from the training data set , wherein , represents an original input vector, represents a historical power output sequence; represents meteorological data; T represents the total number of feature sequences; represents the i-th feature sequence; represents an arbitrary time step;

[0025] S32, inputting the selected feature sequence set into the knowledge guiding layer to extract an implicit feature vector, and then calculating a starting hidden state from the implicit feature vector; represents a starting time step;

[0026] S33, calculating an attention weight of an arbitrary time step pair in the feature sequence set , calculating a historical sequence encoding using the attention weight, aggregating the historical sequence encoding, and extracting a historical state encoding vector ;

[0027] S34, the initial hidden state with the historical state encoding vector is input into the continuous dynamic system modeling layer for fusion to obtain an initial hidden state of a continuous dynamic system, and the continuous dynamic system is constructed based on the initial hidden state of the continuous dynamic system and an ordinary differential equation (ODE), a multi-scale state decomposition mechanism is introduced to rewrite the continuous dynamic system, and the rewritten continuous dynamic system is integrated in a prediction interval to obtain a hidden state sequence at each time in the future ;

[0028] S35, the hidden state sequence at each time in the future is input into a physical perception decoding network to obtain a power prediction value ; the corresponding external condition includes meteorological field forecast data in the future prediction interval

[0029] S36, a total loss function is constructed, and the total loss is calculated according to the power prediction value and the actual power value, and the new energy power prediction network is adjusted in the iteration process to obtain a trained new energy power prediction network.

[0030] Further, the S32 specifically includes the following steps:

[0031] S321, the selected feature sequence set is input into a knowledge guiding layer to extract a hidden feature vector , and the expression is as follows:

[0032] ;

[0033] wherein M represents a pre-trained basic model; represents a parameter of the pre-trained basic model;

[0034] S322, the initial hidden state is calculated according to the hidden feature vector , and the calculation formula is as follows:

[0035] ;

[0036] wherein, , is a learnable linear transformation parameter mapped to the state space of the continuous dynamic system.

[0037] Further, the S33 specifically includes the following steps:

[0038] S331, the feature sequence set ​original input vector at each time step linear projection to a unified hidden space dimension, obtaining an embedding sequence The expression is as follows:

[0039] ;

[0040] wherein, represents a feature embedding, and the calculation formula is as follows:

[0041] ;

[0042] wherein, is a trainable projection matrix, is the dimension of the original input vector ;

[0043] S332, on the embedding sequence , the query vector , the key vector and the value vector at any time step are solved, and the calculation formula is as follows:

[0044] ; wherein,

[0045] , , , respectively represent the query weight, the key weight and the value weight;

[0046] S333, the attention weight at any time step pair is calculated according to the query vector , the key vector and the value vector at any time step , and the calculation formula is as follows:

[0047] ;

[0048] wherein, is the attention mechanism embedding dimension, represents an exponential function with the natural constant e as the base number; represents the i-th time step; represents the key vector at the time step ; is the transpose;

[0049] S334, then the attention weight at any time step pair​ computing an attention output context vector , the calculation formula is as follows:

[0050] ;

[0051] wherein, denotes the value vector of the i-th time step;

[0052] S335, according to the attention output context vector solving the history sequence encoding , the calculation formula is as follows:

[0053] ;

[0054] S336, the history sequence encoding is aggregated, and the end time state value method or the full sequence average pooling method is adopted to extract the history state encoding vector .

[0055] Further, the S34 specifically comprises the following steps:

[0056] S341, fusing the initial hidden state and the history state encoding vector to obtain the initial hidden state of the continuous dynamic system , the calculation formula is as follows:

[0057] ;

[0058] wherein, and are learnable weights and biases, respectively; denotes the vector splicing operation;

[0059] S342, constructing a continuous dynamic system based on the initial hidden state of the continuous dynamic system and the ordinary differential equation ODE, specifically as follows:

[0060] ;

[0061] wherein, denotes the continuous dynamic system; denotes an external factor affecting the future state change of the continuous dynamic system; is a set of learnable parameters of the physical perception decoder;

[0062] S343, introducing a multi-scale state decomposition mechanism, dividing the hidden state sequence of each time to be solved in the future into a fast-changing sub-state and a slow-changing sub-state, and the expression is specifically as follows: ​

[0063] ;

[0064] wherein, represents a fast-changing sub-state; represents a slow-changing sub-state;

[0065] S344, rewriting the continuous dynamic system by using the fast-changing sub-state and the slow-changing sub-state, to obtain a rewritten continuous dynamic system;

[0066] S345, integrating the rewritten continuous dynamic system in a prediction interval to obtain a hidden state sequence , of each time in the future. Hth time step in the future.

[0067] Further, the calculation formula of the power prediction value in the S35 is as follows:

[0068] ;

[0069] wherein, represents a set of learnable parameters inside the physical perception decoding network; represents the physical perception decoding network.

[0070] Further, the total loss function in the S36 is specifically as follows:

[0071] ;

[0072] wherein, L represents the total loss; represents an actual power value; represents a training data set, represents minimizing and optimizing the parameters .

[0073] Another aspect of the present application also provides a new energy power prediction system, which comprises a device end, and the device end adopts the new energy power prediction method to perform power prediction.

[0074] The present application has the following beneficial effects:

[0075] 1. The new energy power prediction method and system provided by the application introduces a continuous dynamic system to overcome the shortcomings of traditional time series prediction models in the face of complex spatiotemporal dynamic evolution. The method can accurately depict the continuous, smooth and reversible dynamic evolution characteristics of the new energy power system, avoiding the error accumulation problem caused by discrete time steps. Through a multi-scale state decomposition mechanism, the new energy power prediction network can effectively capture the dynamic characteristics of rapid disturbances and slow changes, improving the adaptability to different time scales and complex scenarios. In addition, the application adopts a knowledge-guided layer to calculate the initial hidden state , so that the new energy power prediction network has a reasonable initial state that meets the current weather conditions and historical power evolution trend at the initial state, greatly improving the stability and convergence speed of the new energy power prediction network in the training stage. Ultimately, the application can achieve higher accuracy, stability and robustness in new energy power prediction, especially under extreme weather conditions, showing excellent prediction ability and physical consistency, meeting the needs of practical engineering applications.

[0076] 2. The application also discloses a continuous dynamic system, which simulates the dynamic evolution process of the new energy power system by introducing ordinary differential equations (ODEs), avoiding the error accumulation problem caused by the discrete time step method. The new energy power prediction network can effectively simulate the dynamic changes under different time scales (such as short-term weather disturbances and long-term power grid load adjustments) based on the continuous dynamic system, and capture the dynamic characteristics of rapid disturbances and slow changes through a multi-scale decomposition mechanism. In this way, the new energy power prediction network can maintain stability and accuracy during the prediction process, especially when dealing with complex spatiotemporal interactions and extreme weather conditions, it has stronger robustness and prediction ability.

[0077] In addition, the continuous dynamic system also provides a framework for modeling and solving system states in continuous time and space. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 The flowchart of the new energy power prediction method based on the multi-scale state decomposition mechanism in the application;

[0079] Figure 2 The structure diagram of the new energy power prediction network in the application. DETAILED DESCRIPTION

[0080] In order to facilitate the understanding of the application, the application will be described more fully below with reference to the related drawings. The preferred embodiments of the application are shown in the drawings. However, the application can be implemented in many other different forms, and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the application more thorough and comprehensive.

[0081] The current prior art mainly has the following defects in new energy power prediction, multi-modal spatio-temporal sequence modeling and other applications:

[0082] On the one hand, most of the existing models use discrete time step sequence modeling method, which lacks unified dynamic modeling capability of the smooth evolution process of new energy power system in continuous spatio-temporal manifold, and it is difficult to accurately depict its continuous, reversible and smooth complex dynamic evolution characteristics. Especially in complex scenarios such as sudden weather disturbance, rapid power fluctuation and continuous weather field evolution, the long-term prediction stability and trend tracking capability of the model are obviously insufficient. On the other hand, the existing spatio-temporal model generally simply superimposes spatial and temporal information at the feature fusion level, and the spatial feature and time sequence modeling are still decoupled, which cannot effectively express the interactive dynamic process of spatial disturbance diffusion and time accumulation evolution in new energy scenarios, especially in the multi-site, multi-scenario and multi-scale coupling situation, the model has limited ability to capture complex spatio-temporal interaction. In addition, some existing researches try to introduce pre-trained large models to provide prior knowledge of weather field or remote sensing field, but most of them are mainly supplemented by static features, lack of mechanism to deeply integrate large model knowledge into continuous spatio-temporal dynamics learning process, and cannot fully release the cross-scene knowledge transfer and generalization ability contained in the pre-trained model. The present application proposes the following innovative solutions to the above key technical problems in new energy power prediction:

[0083] With reference to Figure 1 The embodiment of the present application provides a new energy power prediction method based on a multi-scale state decomposition mechanism, comprising the following steps:

[0084] S1, collecting multi-source historical data of a wind power station or a photovoltaic power station, preprocessing the multi-source historical data, and then constructing a training data set;

[0085] S2, constructing a new energy power prediction network, the new energy power prediction network comprising a knowledge guiding layer, a historical sequence encoder, a continuous dynamic system modeling layer and a physical perception decoding network; the knowledge guiding layer and the historical sequence encoder are connected in parallel, and the outputs of the knowledge guiding layer and the historical sequence encoder are connected with the continuous dynamic system modeling layer and the physical perception decoding network in sequence;

[0086] S3, training the new energy power prediction network using the training data set to obtain a trained new energy power prediction network;

[0087] S4, deploying the trained new energy power prediction network to a device end and performing power prediction to obtain a prediction result.

[0088] In some embodiments, the S1 specifically comprises the following steps:

[0089] S11, collect multi-source historical data, the multi-source historical data including a historical power output sequence and meteorological data synchronized therewith, if it is a wind power station, the meteorological data including environmental factors affecting the output of the wind turbine; the environmental factors affecting the output of the wind turbine including wind speed (unit: m / s), wind direction (unit: °), etc.

[0090] if it is a photovoltaic power station, the meteorological data including factors affecting the power of the photovoltaic component; the factors affecting the power of the photovoltaic component including solar radiation intensity (unit: W / m²), ambient temperature (unit: °C), etc.

[0091] S12, in the embodiment, all multi-source historical data come from the new energy station data acquisition and monitoring control system, the weather station and the numerical weather prediction; since the multi-source data have different time resolutions and exist noise interference, the multi-source historical data need to be preprocessed, the preprocessing at least including one or more of time alignment, missing value completion and abnormal value elimination, then normalized processing is performed to eliminate the dimension effect, to obtain the preprocessed multi-source historical data;

[0092] S13, a training data set is constructed according to the preprocessed multi-source historical data. The training data set completely reflects the evolution state of the new energy system in the historical period, and is the input of the new energy power prediction network.

[0093] In some embodiments, the historical sequence encoder includes a multi-layer stacked Transformer self-attention encoder, each layer of the Transformer self-attention encoder including a multi-head attention mechanism and a feedforward network, the multi-head attention mechanism and the feedforward network being connected by a residual connection.

[0094] In some embodiments, referring to Figure 2 , the S3 specifically includes the following steps:

[0095] S31, selecting a feature sequence set from the training data set , wherein , represents an original input vector, represents a historical power output sequence; represents meteorological data; represents the total number of feature sequences; represents the i-th feature sequence; represents an arbitrary time step;

[0096] S32, inputting the selected feature sequence set into the knowledge guide layer to extract an implicit feature vector, and then calculating a starting hidden state through the implicit feature vector ; denotes a start time step; denotes a real number set;

[0097] S33, calculating a feature sequence set attention weights of any time step pair, calculating a history sequence encoding by using the attention weights, aggregating the history sequence encoding, and extracting a history state encoding vector ; denotes a real number set;

[0098] S34, inputting the start hidden state and the history state encoding vector into a continuous dynamic system modeling layer for fusion to obtain a start hidden state of a continuous dynamic system, and constructing the continuous dynamic system based on the start hidden state of the continuous dynamic system and an ordinary differential equation (ODE), introducing a multi-scale state decomposition mechanism to rewrite the continuous dynamic system, and integrating the rewritten continuous dynamic system in a prediction interval to obtain a hidden state sequence of each time in the future ;

[0099] S35, inputting the hidden state sequence of each time in the future and corresponding external conditions into a physical perception decoding network to obtain a power prediction value ; wherein the corresponding external conditions may include meteorological field forecast data in the future prediction interval, such as wind speed, irradiance, temperature, etc.

[0100] S36, constructing a total loss function, calculating a total loss according to the power prediction value and an actual power value, iteratively circulating S31 to S35, minimizing the total loss, and adjusting the new energy power prediction network in the process of iteration to obtain a trained new energy power prediction network.

[0101] In some embodiments, the S32 specifically includes the following steps:

[0102] S321, inputting the selected feature sequence set into a knowledge guiding layer to extract a (high-dimensional) hidden feature vector , and the expression is as follows:

[0103] ;

[0104] wherein, denotes a pre-trained basic model; denotes a parameter of the pre-trained basic model; Represents the set of real numbers; This represents the dimension of the output vector of the hidden layer in the model. In this invention, the pre-trained base model has been pre-trained using historical data (i.e., historical power output sequences) before use.

[0105] S322, Based on the latent feature vector Calculate the initial hidden state The calculation formula is as follows:

[0106] ;

[0107] in, , The learnable linear transformation parameters are used to map the state space of a continuous dynamic system. This transformation achieves a smooth projection of the pre-trained knowledge space and the differential manifold space, ensuring the continuity of the model's state space. The knowledge space refers to the knowledge representation or feature space learned in the pre-trained base model. It includes implicit features abstracted from a large amount of historical data or external domain knowledge. By learning from this data, the pre-trained base model gains an understanding of the system's long-term evolution trends, extreme weather patterns, etc., and this knowledge can be transferred and applied to specific tasks. Through knowledge-guided initialization, the new energy power prediction network can integrate this knowledge obtained from the pre-trained large model into new prediction tasks, thereby obtaining a reasonable initial state in the early stages of training and improving system stability and convergence speed.

[0108] Differential manifold space refers to the space in which the state of a system exhibits continuity and reversibility over time. Differential manifold space is used to characterize the dynamic evolution of new energy power systems. For example, the continuous changes of a continuous dynamic system can be described using ordinary differential equations (ODEs), avoiding the error accumulation problem of discrete time steps. This continuous and smooth dynamic evolution process helps improve the stability and accuracy of predictions, especially when facing complex and sudden meteorological disturbances, enabling better capture of the system's long-term trends and local fluctuations.

[0109] The model state space is a multi-dimensional mathematical space that contains all the necessary states within the model at a given moment (or time step). The state space includes not only historical information about the power system, but also the dynamic evolution process modeled through differential equations and knowledge gained from pre-trained models.

[0110] Through this knowledge-guided initialization mechanism, the model has a reasonable initial state that conforms to the current actual working conditions at the beginning of the dynamic system solution. This avoids the problems of system divergence, oscillation and long convergence caused by initial guesses without physical basis in the early training stage, and greatly improves the overall stability, interpretability and transfer generalization ability of continuous dynamic system modeling.

[0111] The present application aims at the problem that the historical observation information of new energy power system is limited and the model initialization is difficult in different scenarios, and proposes a knowledge guiding layer. Through pre-training on large-scale cross-scenario new energy system historical data, the new energy power prediction network has the learning ability of long-term evolution trend, complex disturbance characteristics and extreme weather mode. In the specific prediction task, the pre-training model is used to encode the current observation history sequence, and directly output the hidden feature vector representing the current system physical state as the initial state input of the continuous dynamics model. Compared with the traditional random or zero vector initialization method, the present application can make the model have a reasonable starting state conforming to the current weather condition and the historical power evolution trend from the beginning, which significantly improves the stability, convergence speed and robustness of the new energy power prediction network under extreme weather conditions in the early training stage.

[0112] In some embodiments, the S33 specifically comprises the following steps:

[0113] S331, the feature sequence set is linearly projected to map to a unified hidden space dimension to obtain an embedding sequence , the expression is specifically as follows:

[0114] ;

[0115] wherein, represents the feature embedding, and the calculation formula is specifically as follows:

[0116] ;

[0117] wherein, is a trainable projection matrix, is the dimension of the original input vector ;

[0118] S332, on the embedding sequence , for any time step , the query vector , the key vector and the value vector of any time step are solved, and the calculation formula is as follows:

[0119] ;

[0120] wherein, , , represent the query weight, the key weight and the value weight respectively;

[0121] ​S333, the attention weight of any time step t is calculated according to the query vector Q, the key vector K and the value vector V, the calculation formula is as follows:

[0122]

[0123] , where, is the embedding dimension of the attention mechanism, which is used to avoid calculation instability caused by too large numerical value, is the exponential function with the natural constant e as the base number; is the transpose;

[0124] S334, then the attention output context vector is calculated according to the attention weight of any time step t, the calculation formula is as follows:

[0125]

[0126]

[0127] S335, the history sequence encoding is solved according to the attention output context vector, the calculation formula is as follows:

[0128]

[0129] S336, the history sequence encoding is aggregated, and the last time state value method or the full sequence average pooling method is used to extract the history state encoding vector

[0130] In some embodiments, the last time state value method and the full sequence average pooling method in S336 are as follows:

[0131] Last time state value method (directly taking the encoding vector of the last time of the history sequence):

[0132]

[0133] ​​​​​​​​​​​​​​​​​​​​​​​​​​Full sequence average pooling method (taking the average of all time encoding vectors to aggregate):

[0134] ;

[0135] wherein, The evolution trend, disturbance change and dynamic characteristics of the new energy system during the history period are summarized in full.

[0136] In some embodiments, the S34 specifically comprises the following steps:

[0137] S341, fusing the initial hidden state and the historical state encoding vector to obtain the initial hidden state of the continuous dynamic system , and the calculation formula is as follows:

[0138] ;

[0139] wherein, and are learnable weights and biases, respectively; denotes a vector splicing operation;

[0140] S342, constructing a continuous dynamic system based on the initial hidden state of the continuous dynamic system and the ordinary differential equation ODE, specifically as follows:

[0141] ;

[0142] wherein, denotes the continuous dynamic system; in order to ensure the continuity, smoothness and reversibility of the dynamic system evolution process in the state space, the continuous dynamic system satisfies the differential homeomorphism characteristic, that is, the state flow field transformation of the continuous dynamic system remains reversible and has no singular point in the local space; denotes external factors affecting the future state change of the continuous dynamic system; the external factors are external conditions or environmental variables of the continuous dynamic system, which are used to determine the change of the system state at a future time in the prediction process. In new energy power prediction, the external factors include meteorological factors (such as wind speed, solar irradiance, temperature, etc.) and other external factors that may affect the power output of new energy stations (such as wind farms or photovoltaic farms). These external factors are necessary inputs for the new energy power prediction network when predicting future power output, because they directly affect the state change and dynamic evolution process of the continuous dynamic system. denotes a set of learnable parameters of the physical perception decoder; denotes the hidden state to be solved;

[0143] S343, introduce a multi-scale state decomposition mechanism, divide the hidden state sequence of each time to be solved in the future into a fast-changing substate and a slow-changing substate, and the expression is specifically as follows:

[0144] ;

[0145] Among them, indicates the fast-changing substate; indicates the slow-changing substate;

[0146] S344, rewriting the continuous dynamic system by using the fast-changing substate and the slow-changing substate, to obtain a rewritten continuous dynamic system;

[0147] S345, integrating the rewritten continuous dynamic system in the prediction interval to obtain the hidden state sequence , , wherein Hth time step in the future. The integration process adopts a high-order numerical integration algorithm with adaptive step size (such as Runge-Kutta-Fehlberg method), which ensures high precision stability while avoiding the problem of time granularity dependence in traditional discrete recursive methods.

[0148] In view of the characteristics of new energy power system, such as continuity, reversibility, multi-time scale dynamic coupling and the like, the traditional prediction model based on discrete recursion has problems of error accumulation, insufficient long-period prediction accuracy and the like. On the basis, the present application introduces a continuous dynamic system constructed by ordinary differential equation ODE, and converts the dynamic evolution process of the new energy power system into the solution of the ordinary differential equation ODE in the continuous time domain, thereby avoiding the cumulative error problem caused by the discrete time step. At the same time, the present application proposes a multi-scale state decomposition mechanism, which models the fast disturbance (such as wind array, cloud shielding) and slow trend (such as day-night cycle, power grid scheduling load) of the new energy power system respectively, effectively improves the adaptability of the model to different time scale dynamic characteristics, and enhances the dynamic capturing ability and long-term prediction stability in complex scenarios.

[0149] In addition, the present application improves the unified modeling and prediction accuracy of the new energy power prediction network for high-frequency disturbance, short-term drastic change and long-term trend through independent modeling and coupled solution mechanism.

[0150] Independent modeling refers to introducing a multi-scale state decomposition mechanism to divide the dynamic change of the new energy power system into a fast-changing substate (such as fast disturbance caused by sudden weather events such as wind array change and cloud shielding) and a slow-changing substate, so that the new energy power prediction network can process the evolution process of the new energy power system at different time scales;

[0151] The coupling solution mechanism refers to how to unify and couple the solutions of the sub-states after independent modeling of different sub-states, so as to obtain the overall evolution trajectory of the new energy power system. In the present application, ordinary differential equations (ODE) are used to describe the evolution of the state of the new energy power system, and a multi-scale state decomposition mechanism is introduced. This mechanism enables the new energy power prediction network to handle fast-changing and slow-changing states separately, but ultimately solves the state at each time in the new energy power system through coupling. This ensures that the long-term trend (i.e., the slow-changing sub-state) and the short-term disturbance (i.e., the fast-changing sub-state) of the new energy power system can be effectively modeled and predicted in the same prediction process.

[0152] In some embodiments, the power prediction value in S35 is calculated as follows:

[0153] ;

[0154] wherein, represents a set of learnable parameters inside the physical perception decoding network; represents the physical perception decoding network.

[0155] The physical perception decoding network fully incorporates the physical constraints of new energy power in design. For example, in the wind power scenario, the output of the physical perception decoding network is limited within the range of the rated power of the wind turbine, and the physical law of the wind turbine power curve is combined to realize nonlinear mapping; in the photovoltaic scenario, the physical perception decoding network ensures that the output does not exceed the array capacity, and automatically returns to zero under specific conditions such as at night.

[0156] In the state solving process of the continuous dynamic system, the present application uses a high-order numerical integration algorithm to realize dynamic control of integration error and stable advancement of state flow evolution trajectory. By dynamically evaluating the state evolution speed in each integration small step, the time step is automatically adjusted, taking into account the transient response ability of high-frequency disturbance and the long-term maintenance ability of low-frequency smooth trend, thereby ensuring the stability, physical continuity and reversibility of the entire state flow field within the prediction interval, and improving the global convergence and engineering application safety of the new energy power prediction network.

[0157] In some embodiments, the total loss function in S36 is specifically as follows:

[0158] ;

[0159] wherein, L represents the total loss; represents the actual power value; represents the training data set;​​​​ representing a parameter is minimized.

[0160] The following is described by way of example for ease of understanding:

[0161] 1) Application background of new energy power prediction

[0162] The disclosed new energy power prediction method based on a multi-scale state decomposition mechanism has been tested in actual new energy power prediction scenarios, covering two typical new energy scenarios of wind farms and photovoltaic power stations. The test task covers short-term rolling prediction and intra-day continuous prediction business requirements, and the prediction time scale varies from 10 minutes to 24 hours in the future. During the application, the new energy power prediction network is based on the Supervisory Control And Data Acquisition (SCADA) monitoring data, real-time weather observation data and future weather numerical prediction data to carry out modeling and prediction, fully reflecting the complex characteristics of the dynamic changes of actual new energy power generation output with weather conditions.

[0163] 2) Data sources and modeling process

[0164] In the training and testing process of the new energy power prediction network, the data used includes new energy station measured historical power data, weather observation data (wind speed, wind direction, irradiance, temperature, humidity, etc.) and high-precision weather numerical prediction data. The historical data period covers multiple complete climate seasons, and the sampling time interval is uniform at 10 minutes. The pre-trained basic model is first pre-trained using long-term historical data (i.e. historical power output sequence), to learn the universal evolution law of the new energy power system, and then combined with the historical observation data of the target station for transfer fine-tuning training, to adapt to the power evolution characteristics in specific scenarios. During the training process, an adaptive learning rate optimization algorithm is used, and the overall training process of the pre-trained basic model is stable and convergent, and the training time is controllable.

[0165] 3) Model structure and deployment configuration

[0166] The new energy power prediction network adopts a complete multi-module collaborative architecture in deployment and application. The input end completes multi-source data alignment, outlier elimination and normalization processing through the data preprocessing module; the knowledge guided layer uses the pre-trained large model to generate the initial hidden state of the system; the historical sequence encoder extracts dynamic dependency features in the historical observation sequence through the Transformer; the physical perception decoding network fuses the hidden state and future weather forecast information to output a predicted power sequence with physical consistency. The overall calculation process of the new energy power prediction network is clear, the single prediction calculation time is short, and it meets the real-time requirements of the on-site engineering dispatching system.

[0167] 4) Forecast application scenario examples;

[0168] In the wind power application scenario, the new energy power prediction network in the application can continuously predict the wind power output change process in the next 24 hours based on future weather forecasts and current station power status. The prediction results can be used for power grid safety scheduling, load balancing and reserve capacity arrangement. In the photovoltaic scenario, the new energy power prediction network can respond to future irradiance and weather disturbance changes in real time, dynamically output future hourly photovoltaic power generation output, and provide prediction support for intraday power balance and new energy consumption. In the scenarios of extreme weather changes, rapid weather disturbances and drastic changes in morning and evening sunlight, the new energy power prediction network can still continuously and stably output power generation prediction curves that conform to physical logic. The prediction results have good stability and continuity, and have strong engineering application applicability.

[0169] Another aspect of the application also provides a new energy power prediction system, which comprises a device end that adopts the new energy power prediction method to perform power prediction.

[0170] The above is only a specific embodiment of the application, but the protection scope of the application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered within the protection scope of the application. Moreover, the technical solutions of each embodiment of the application can be combined with each other, but it must be based on the realization by a person skilled in the art. When the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist and is not within the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A new energy power prediction method based on a multi-scale state decomposition mechanism, characterized in that, Includes the following steps: S1. Collect multi-source historical data from wind farms or photovoltaic power stations, preprocess the multi-source historical data, and then construct a training dataset. S2. Construct a new energy power prediction network, which includes a knowledge guidance layer, a historical sequence encoder, a continuous dynamic system modeling layer, and a physical perception decoding network. The knowledge guidance layer and the historical sequence encoder are connected in parallel, and their outputs are connected to the continuous dynamic system modeling layer and the physical perception decoding network in sequence. S3. Train the new energy power prediction network using the training dataset to obtain the trained new energy power prediction network. S4. Deploy the trained new energy power prediction network to the device and perform power prediction to obtain the prediction results; S3 specifically includes the following steps: S31. Select a set of feature sequences from the training dataset. ,in , Represents the original input vector. Represents the historical power output sequence; Represents meteorological data; T Indicates the total number of characteristic sequences; Represents the i-th feature sequence; Represents any time step; S32. Select the set of feature sequences The input is fed into the knowledge guidance layer, where hidden feature vectors are extracted. The initial hidden state is then calculated using these hidden feature vectors. ; Indicates the starting time step; S33, Calculate the set of feature sequences Attention weights are assigned to any pair of time steps within the time frame. These attention weights are then used to calculate the historical sequence encoding. The historical sequence encodings are then aggregated to extract the historical state encoding vector. ; S34, Set the initial hidden state With historical state encoding vector The inputs are fused into the continuous dynamic system modeling layer to obtain the initial hidden state of the continuous dynamic system. Based on the initial hidden state and the differential equation ODE, the continuous dynamic system is constructed. A multi-scale state decomposition mechanism is introduced to rewrite the continuous dynamic system. The rewritten continuous dynamic system is then integrated over the prediction interval to obtain the hidden state sequence for each future time step. ; S35. Sequence of hidden states at each future time step. With corresponding external conditions The input is fed into the physical sensing decoding network to obtain the power prediction value. ; Corresponding external conditions Including weather forecast data for the future forecast period; S36. Construct the total loss function based on the predicted power values. The total loss is calculated based on the actual power value. The process is iterated from S31 to S35 to minimize the total loss. During the iteration, the new energy power prediction network is adjusted to obtain the trained new energy power prediction network. S32 specifically includes the following steps: S321. Select the set of feature sequences. The input is fed into the knowledge guidance layer to extract the latent feature vector. The expression is as follows: ； Where M represents the pre-trained base model; Represents the parameters of the pre-trained base model; S322, Based on the latent feature vector Calculate the initial hidden state The calculation formula is as follows: ； in, , These are the learnable linear transformation parameters that map to the state space of a continuous dynamic system.

2. The new energy power prediction method based on multi-scale state decomposition mechanism according to claim 1, characterized in that, S1 specifically includes the following steps: S11. Collect multi-source historical data, which includes historical power output sequences and synchronous meteorological data. If it is a wind farm, the meteorological data includes environmental factors that affect the wind turbine output. For photovoltaic power plants, meteorological data includes factors that affect the power output of photovoltaic modules; S12. Preprocess the multi-source historical data. The preprocessing includes at least one or more of the following: time alignment, missing value completion, and outlier removal. Then, normalize the data to obtain the preprocessed multi-source historical data. S13. A training dataset is constructed based on the preprocessed multi-source historical data.

3. The new energy power prediction method based on multi-scale state decomposition mechanism according to claim 2, characterized in that, The historical sequence encoder includes multiple stacked Transformer self-attention encoders. Each Transformer self-attention encoder includes a multi-head attention mechanism and a feedforward network, which are connected by residuals.

4. The new energy power prediction method based on multi-scale state decomposition mechanism according to claim 3, characterized in that, S33 specifically includes the following steps: S331, Set the feature sequences The original input vector at each time step Perform a linear projection to map to a unified hidden space dimension to obtain the embedding sequence. The specific expression is as follows: ； in, The feature embedding is represented by the following formula: ； in, For trainable projection matrices, The original input vector The dimension; S332, in the embedded sequence Above, for any time step Solve for any time step query vector Key vector Sum value vector The calculation formulas are as follows: ； in, , , These represent query weight, key weight, and value weight, respectively. S333, based on any time step query vector Key vector Sum value vector Calculate arbitrary time step pairs attention weights The specific calculation formula is as follows: ； in, Embedding dimensions for attention mechanisms This represents an exponential function with the natural constant e as the base. Indicates the first One time step; Indicates time step The key vector; For transpose; S334, then based on arbitrary time steps attention weights Calculate the attention output context vector The calculation formula is as follows: ； in, Indicates the first A vector of values ​​for each time step; S335. Output context vector based on attention. Solving historical sequence encoding The calculation formula is as follows: ； S336. Encoding historical sequences Aggregation is performed, and the historical state encoding vector is extracted using either the final time-state value method or the full sequence average pooling method. .

5. The new energy power prediction method based on multi-scale state decomposition mechanism according to claim 4, characterized in that, S34 specifically includes the following steps: S341, Change the initial hidden state With historical state encoding vector By merging the states, the initial hidden states of the continuous dynamic system can be obtained. The calculation formula is as follows: ； in, and These are the learnable weights and biases, respectively. This represents a vector concatenation operation; S342, Initial Hidden State Based on Continuous Dynamic Systems And the construction of continuous dynamic systems using ordinary differential equations (ODEs), as detailed below: ； in, Represents a continuous dynamic system; These represent external factors that influence the future state changes of a continuous dynamic system. This is the set of learnable parameters for the physical perception decoder; S343. Introduce a multi-scale state decomposition mechanism to divide the hidden state sequence at each future time step into rapidly changing sub-states and slowly changing sub-states. The specific expression is as follows: ； in, Indicates a rapidly changing substate; Indicates a slowly changing substate; S344. Rewrite the continuous dynamic system using fast-changing sub-states and slow-changing sub-states to obtain the rewritten continuous dynamic system. S345. Utilizing a rewritten continuous dynamic system in the prediction interval Integrating within the inner timeframe yields the sequence of hidden states at each future time step. , ,in The Hth time step in the future.

6. The new energy power prediction method based on multi-scale state decomposition mechanism according to claim 5, characterized in that, The power prediction value in S35 The calculation formula is as follows: ； in, This represents the set of learnable parameters within the physical sensing decoding network; This represents a physical sensing decoding network.

7. The new energy power prediction method based on multi-scale state decomposition mechanism according to claim 6, characterized in that, The total loss function in S36 is as follows: ； Where L represents the total loss; Indicates the actual power value; Represents the training dataset. Indicates the parameter Perform a minimization optimization.

8. A new energy power prediction system, characterized in that, The device side uses the new energy power prediction method described in any one of claims 1 to 7 to perform power prediction.

Images