A photovoltaic power early warning method and system based on diffusion model and hidden space

Abstract

This invention provides a photovoltaic power early warning method and system based on a diffusion model and latent space. The method first preprocesses multi-source data such as irradiance, temperature, and equipment status, and then decouples these data into trend, disturbance, and state features through feature decoupling. Subsequently, a dual-model coupling framework is constructed: multi-scenario probabilistic paths for generating future power are used to characterize its randomness and multimodal characteristics; simultaneously, a latent space temporal model is employed to depict the evolution of potential states, and diffusion output constrains state transitions, achieving collaborative modeling. Furthermore, uncertainty is statistically quantified through multi-path analysis, and concept drift is identified by combining residual analysis and change point detection, analyzing its causes and intensity. Finally, uncertainty and drift information are integrated to construct a risk score, achieving tiered early warning and effectively improving the accuracy, robustness, and interpretability of the early warning.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation prediction and intelligent operation and maintenance technology, and in particular relates to a photovoltaic power early warning method and system based on diffusion model and latent space. Background Technology

[0002] Photovoltaic power generation is influenced by multiple factors, including irradiance, temperature, module temperature, cloud cover, wind speed, equipment status, power curtailment control, and grid dispatch, exhibiting randomness, volatility, coupling, and non-stationarity. In particular, under scenarios such as rapid cloud movement, localized cloud cover, significant equipment operating condition switching, power curtailment, and fault disturbances, power output often shows significant fluctuations and multi-stage variations.

[0003] Existing methods are mostly deterministic or simple interval predictions, which are effective under stable conditions, but have shortcomings: they are difficult to express multimodal distributions and cannot reflect the diversity of scenarios under complex weather conditions; they lack modeling of the dynamic evolution of power and are difficult to reveal the state transition mechanism; they have insufficient ability to identify concept drift and cannot accurately locate the triggering time and intensity; and they lack interpretability and are difficult to identify key triggers.

[0004] Therefore, single probabilistic prediction models struggle to simultaneously characterize short-term random disturbances and continuous state evolution, while single state evolution models fail to fully express the multimodal distribution characteristics under complex weather conditions. There is an urgent need to propose a photovoltaic power early warning method that integrates probabilistic scenario generation, dynamic evolution modeling, uncertainty quantification, concept drift detection, and trigger identification to improve the accuracy, interpretability, and robustness of early warnings. Summary of the Invention

[0005] Purpose of the invention: The technical problem to be solved by the present invention is to provide a photovoltaic power early warning method and system based on diffusion model and latent space to address the shortcomings of the prior art. This solves the problem that the prior art is difficult to simultaneously take into account the expression of multimodal uncertainty, conceptual identification and cause identification, thereby achieving predictable, quantifiable and explainable early warning of photovoltaic power generation operation risks.

[0006] The method includes the following steps:

[0007] Step 1: Obtain meteorological data, electrical operation data, and operation and maintenance status data from multi-source data related to photovoltaic power generation. Perform time alignment, missing value repair, anomaly removal, and standardization on the multi-source data. Based on feature volatility, time variability, and power correlation, construct a feature adaptation discriminant function to decouple multi-source features into trend-driven features, random disturbance features, and operation status features, forming a model adaptation input representation.

[0008] Step 2: Select random perturbation features and operating state features as conditional inputs, and use the forward noise addition and reverse noise reduction process to generate a multi-scenario probability path for future power in order to capture the uncertainty and multimodal features of future power.

[0009] Step 3: Select trend-driven features and operating state features as inputs to the latent space-time model, and combine the output multi-scenario samples as uncertainty references. The encoder maps the samples to the latent state space to establish latent state transition relationships, and the decoder generates future latent state trajectories and corresponding power trends.

[0010] Step 4: Merge the generated future power probability path with the power path decoded by the latent space-time model to form a joint power path sample set; perform statistical analysis on the joint sample set to calculate the sample mean, variance, quantiles, confidence intervals and interval widths, and quantify the range of future power fluctuations and prediction uncertainty.

[0011] Step 5, Concept Drift Detection: By detecting the changing trend of the data, the triggering factor is located, and then the change point detection technology is further applied to output the drift triggering period and the corresponding time position.

[0012] Step 6, Feature contribution analysis and drift cause identification: Locate key causes and drift intensity by detecting changes in feature contribution;

[0013] Step 7, Risk Assessment and Tiered Early Warning: Calculate a comprehensive risk score by combining uncertainty indicators and final drift intensity.

[0014] Step 1 includes the following steps:

[0015] Step 1.1: Collect data from multiple sources;

[0016] Collect multi-source time-series data related to photovoltaic power prediction and early warning, including meteorological data, electrical operation data, and operation and maintenance status data;

[0017] The meteorological data includes irradiance, ambient temperature, component temperature, wind speed, cloud cover, humidity, and air pressure.

[0018] The electrical operating data includes voltage, current, active power, reactive power, and inverter operating status parameters.

[0019] The operation and maintenance status data includes start / stop status, protection action, blocking status, cleaning status, limited generation status, and control commands;

[0020] Let multi-source observation data at time t be used. for:

[0021] ;

[0022] in, This represents the vector of environmental variables at time t. This represents the power and electrical quantity vector at time t. This represents the device and its operational status vector at time t.

[0023] Step 1.2, Alignment and Standardization;

[0024] Perform unified time alignment, missing value repair, outlier removal, and normalization on multi-source data. The input at time t is then standardized. Represented as:

[0025] ;

[0026] in, and These represent the mean and standard deviation obtained from the training samples, respectively.

[0027] Constructing the historical input sequence at time t using a sliding time window :

[0028] ;

[0029] The corresponding future target prediction sequence is represented as follows:

[0030] ;

[0031] in, It is the length of the historical window. It is the future predicted step size. Indicated by time The future as a starting point Step photovoltaic power output sequence; Indicates the predicted time Photovoltaic power output value at the location, ;

[0032] This step outputs a standardized historical input sequence. and future target sequence and the historical input sequence Step 1.3 involves decoupling and recombining model-adaptive features.

[0033] Step 1.3, Decoupling and Recombination of Model-Adaptive Features;

[0034] Obtaining historical input sequences Subsequently, in order to adapt the multi-source input features to the modeling requirements of random perturbations and multimodal distributions, as well as the modeling requirements of the latent space-time model for continuous evolution states, model-adaptive feature decoupling and recombination processing is performed on each input feature in the historical input sequence.

[0035] Let the historical input sequence be... The Middle The characteristic variables are First, calculate the volatility index of this feature within the historical window. :

[0036] ;

[0037] in, Indicates the first The average value of each feature variable within a historical window.

[0038] Calculate the first Time change indicators of each characteristic variable :

[0039] ;

[0040] in, It is used to characterize the magnitude of change of this feature between adjacent time points.

[0041] Calculate the first Time-related indicators between characteristic variables and power output :

[0042] ;

[0043] in, This represents the function for calculating the correlation coefficient. Indicates time Photovoltaic power output.

[0044] Constructing a feature-adaptive discriminant function :

[0045] ;

[0046] in, These are non-negative weighting coefficients. The larger the value, the more random perturbation the feature exhibits; The smaller the value, the more trend-driven the feature is.

[0047] Based on the discrimination threshold The input features are divided into trend-driven feature sets. and random perturbation feature set :

[0048] ;

[0049] Simultaneously, discrete state variables such as equipment operating status, limited-emission status, protection actions, and control commands are constructed into operating status codes. And generate a model-adapted input representation:

[0050] ;

[0051] in, The model adapts the input representation, which serves as the joint input to the subsequent latent space-time model.

[0052] in, Used as input to the latent space-time model to characterize the potential state evolution process of photovoltaic power; The conditions used are to characterize the random fluctuations and multimodal uncertainties of future power. State-related variables such as equipment operating status, power limiting status, protection actions, and control commands are not included in the above two types of continuous feature classification, but are instead encoded separately as operating status codes. , as a shared constraint input with the latent space-time model.

[0053] Through the above feature decoupling and recombination process, trend-driven features are mainly used for latent state modeling, random perturbation features are mainly used for probabilistic scenario generation, and running state features are used as shared constraint information to participate in the modeling process of the two models. This improves the collaborative modeling capability with the latent space-time model and avoids the problem of insufficient model adaptability caused by direct mixing of different types of features as input.

[0054] Step 2 includes the following steps:

[0055] Step 2.1: Generate future power probability scenarios;

[0056] At any moment Model adaptation input representation In the selection of random perturbation feature set and running status coding As a conditional input, learn the future Photovoltaic power output sequence Conditional probability distribution:

[0057] ;

[0058] in, For parameters, This represents a conditional probability distribution as defined.

[0059] To characterize the multimodal features of future power output, a diffusion probability modeling approach is adopted. The future power distribution is learned through progressive noise addition and denoising. The forward diffusion process is represented as:

[0060] ;

[0061] in, Represented as in the known number Step state Under the conditions, the first Step state The conditional probability distribution; For the first Step-diffusion noise figure, It is the identity matrix. Indicates a Gaussian normal distribution;

[0062] Equivalent reparameterization is represented as:

[0063] ;

[0064] in, Indicates the sequence from the initial sequence to the nth. The proportion of signal components retained after step diffusion. Indicates random noise; This can be represented as raw data (initial state) or a target sequence without noise.

[0065] The inverse denoising process restores the target distribution under given historical conditions; the loss function (training objective function) of the diffusion model. Represented as:

[0066] ;

[0067] in, This represents a denoising network. This represents the mathematical expectation operator; this loss function is typically used to constrain the prediction error of the denoising network on noise or target data, enabling the model to progressively learn the mapping relationship from noise to the true data distribution.

[0068] To express the temporal dependencies between future multi-step outputs, the joint conditional distribution of future power is recursively expanded over the prediction step:

[0069] ;

[0070] in, Represented as from time At the time A sequence of photovoltaic power over a continuous time period;

[0071] Two or more future power generation scenarios are obtained through repeated sampling, at any given time. The generated future power path sample set for:

[0072] ;

[0073] in, Indicates the first A future power path, The value ranges from 1 to N, where N represents the total number of sample paths;

[0074] The set of scenarios output by this step represents the various possible evolutionary outcomes of future power under the same historical conditions, providing an external stochastic reference for the state evolution modeling of the subsequent latent space-time model, and providing a sample basis for uncertainty quantification.

[0075] Step 3 includes the following steps:

[0076] Step 3.1, construct the latent space-time model;

[0077] A latent space-time model is constructed based on the model-adapted input representation. This model is then used to describe the dynamic relationship between environmental driving forces and power evolution from the perspective of latent states. The generated multi-scenario probability paths are used to characterize the random perturbation boundary of future power and can serve as consistency reference information for the latent state transition or decoding output of the latent space-time model.

[0078] First, the trend-driven feature set is encoded using an encoder. and running status coding Mapping to the latent state space:

[0079] ;

[0080] in, For a moment The potential state representation, For encoder functions, These are encoder parameters.

[0081] The potential state transitions are modeled in a probabilistic form to characterize the stochasticity of the future process:

[0082] ;

[0083] in, This indicates the central tendency of potential state transitions. This represents the uncertainty of the potential state at that moment; it is defined as:

[0084] ,

[0085] ;

[0086] in, Represents the mean mapping function, Represents the covariance mapping function. This represents the external driving vector, which encompasses meteorological environmental variables, power grid dispatching commands, and equipment control signals.

[0087] The decoder maps the latent states back to the power output space at time 1000. Photovoltaic power forecast for:

[0088] ;

[0089] in, Represents the decoder function;

[0090] This step outputs the future potential state trajectory and the corresponding power path, which, together with the generated future power probability path, serve as the input to the uncertainty quantification module in step 4.

[0091] Step 4 includes the following steps:

[0092] Step 4.1, Multi-scenario Statistics;

[0093] Suppose that at a given time, the path generated and the output path of the latent space time model are obtained together. Given a sample of future power paths, the sample mean at any prediction time is... for:

[0094] ;

[0095] in, Indicates the first The future power path at the prediction time Power value at;

[0096] Step 4.2, uncertainty quantification;

[0097] At any moment Place, Variance of power path samples for:

[0098] ;

[0099] in, Used to quantify the magnitude of uncertainty in the prediction result at that moment;

[0100] At the predicted time At the point, the empirical distribution of the first quantiles Defined as:

[0101] ;

[0102] in, Indicates the quantile probability level. Indicates by The empirical distribution function constructed from the sample paths, Represented as a power value variable; Indicates the infimum operation;

[0103] Construct confidence intervals:

[0104] ;

[0105] in, Represented as time The confidence interval, Represented as time The 5th percentile (lower quantile); Represented as time The 95th percentile (upper quantile value); Represented as time The width of the interval (the width of uncertainty);

[0106] The interval width is defined as:

[0107] ;

[0108] The wider the interval, the more uncertain the future power. This serves as input for step 5, concept drift detection, and step 7, risk assessment.

[0109] Step 5 includes the following steps:

[0110] Step 5.1, Concept drift detection;

[0111] To identify changes in model applicability conditions and shifts in system operating mechanisms, this paper uses prediction residuals, distribution shifts, and change point detection to identify concept drift; the prediction residual is defined as:

[0112] ;

[0113] in, For a moment The predicted residuals Indicates time The actual observed power value, This represents the time obtained from the joint output with the latent space-time model. The predicted power value.

[0114] When the mean, variance, or local statistical behavior of the residuals consistently deviates from the reference level, the system can be considered to have experienced a drift trend. Furthermore, by comparing the input and output distributions of the reference window and the current window, distribution shift analysis can be performed using information divergence, transmission distance, or sample distribution significance tests to determine the degree of difference between historical and current operating conditions in terms of input and output distributions. If the current window deviates significantly in both residual statistics and distribution shift, it is considered a high-confidence period of conceptual drift.

[0115] To enhance the ability to pinpoint the trigger time, a change point detection method can be used to identify cumulative changes in residuals and inflection points in local means. This step outputs the drift trigger period and its corresponding time position, which serves as input for drift cause identification in step 6.

[0116] Step 6 includes the following steps:

[0117] Step 6.1, Feature contribution analysis;

[0118] After detecting concept drift, we further analyze the contribution of trend-driven features, random disturbance features, and operational status features to changes in prediction bias or uncertainty, thereby identifying the key causes of drift. Let the set of variable nodes V be: ;

[0119] in Indicates the first The feature variables include irradiance, temperature, cloud cover, wind speed, voltage, current, inverter status, protection actions, power limiting commands, and power output; n represents the total number of feature variables.

[0120] The reference window is selected as the stable operating window before the drift trigger (such as the past 24 hours), and the current window is the recent window containing the trigger time (such as the past 1 hour).

[0121] Based on the current drift window For samples within the range, each feature is calculated using model interpretation methods (e.g., SHAP value, feature gradient contribution, or local perturbation analysis). Contribution to model prediction residuals or uncertainty indicators Define the change in feature contribution relative to a reference window (such as the pre-drift stable operating window). for:

[0122] ;

[0123] in The average contribution of the same feature under the reference window; Sort by largest to smallest, and select the top... These features serve as a set of key inducing variables. ;

[0124] Step 6.2, Drift Cause Analysis and Intensity Quantification;

[0125] Definition of the first Drift intensity index of each window for:

[0126] ;

[0127] in, This indicates the degree of deviation of key factors from the reference operating condition (e.g., the standardized deviation of eigenvalues). and These are the weighting coefficients. The first term measures the overall magnitude of change in the feature contribution, and the second term measures the degree of anomalousness in the values ​​of the key factors themselves.

[0128] Based on the characteristic category of the key inducing variables, drift inducing factors can be divided into trend evolution inducing factors, random disturbance inducing factors, and operational state constraint inducing factors.

[0129] In practical implementation, the calculation of feature contributions can employ model-independent interpretation methods or model-dependent gradient or attention mechanisms. For example, under normal clear-sky conditions, irradiance typically contributes the most; under cloud shadow disturbance conditions, the contributions of cloud cover and irradiance increase significantly; under equipment thermal derating conditions, the contribution of component temperature or inverter temperature rise increases; and under control-constrained conditions, the contribution of limiting command becomes dominant. By calculating the feature contribution distribution of each window through a sliding window, it is possible to observe which features' contributions change significantly before and after drift.

[0130] The identification of key contributing variables is based on the ranking of changes in contribution; the larger the absolute value of the change, the stronger the influence of the feature on the current drift. This step outputs key contributing variables, a description of changes in feature contribution, and a drift intensity index, which serve as inputs for the risk assessment and graded early warning module.

[0131] Step 7 includes the following steps:

[0132] Step 7.1, Risk Assessment;

[0133] After obtaining uncertainty indicators and drift intensity, this invention performs a joint assessment of multi-source risk information to form a unified risk score;

[0134] The risk score comprehensively considers the volatility of the future power prediction distribution, the width of the quantile interval, the drift intensity, the real-time residual amplitude, and the degree of anomaly of key factors, and is expressed as follows:

[0135] ;

[0136] in, For a moment Comprehensive risk score; Indicates time Prediction standard deviation (characterizing the degree of dispersion); Indicates time Confidence interval width (characterizing the range of uncertainty); Indicates time Drift intensity index (reflects changes in feature contribution and the degree of deviation of key factors); Indicates time The intensity of key anomalies (obtained by weighting the change in contribution or deviation of eigenvalues ​​of key inducing variables), which can be obtained by weighting the change in contribution and deviation of eigenvalues ​​of the key inducing variables identified in step 6; These are non-negative weighting coefficients used to adjust the contribution of each risk component to the overall score;

[0137] Step 7.2, tiered early warning;

[0138] Based on different power plants, seasons, and operating conditions, the weight coefficients and scoring thresholds can be calibrated according to historical statistical results or operating experience, and the risks can be classified into low risk, general risk, relatively high risk, and high risk levels.

[0139] During the early warning output phase, this invention does not only output a single-level label, but also simultaneously outputs the potential future power fluctuation range, risk level, main inducing variables, description of changes in characteristic contributions, and corresponding handling suggestions. Differentiated response strategies can be generated for different inducing factor types.

[0140] The present invention also provides a photovoltaic power early warning system based on a diffusion model and latent space implemented using the aforementioned method, comprising:

[0141] The multi-source data acquisition, alignment, and standardized modeling module is used for cleaning and standardizing multi-source data and initiating a historical project input window.

[0142] The model-adaptive feature decoupling and recombination module is used to divide historical input features into trend-driven features, random disturbance features, and operating state features based on feature volatility, time variability, and power correlation, and generate model-adaptive input representations.

[0143] The future power probability scenario generation module is used to generate future power multimodal probability scenario samples based on random perturbation features and operating state encoding.

[0144] The latent space-time model construction module is used to map trend-driven features and operating state codes to the latent state space to characterize the potential state evolution process of photovoltaic power.

[0145] The dual-model path fusion and uncertainty quantification module is used to perform joint statistical analysis on multi-path samples output by the implicit space-time model to quantify the range of uncertainty in future predictions.

[0146] The concept drift detection module is used to identify concept triggers based on the joint analysis of residuals and distribution shifts.

[0147] The feature contribution analysis and drift cause identification module is used to detect and locate key causes and drift intensity through changes in feature contribution.

[0148] The risk assessment and tiered early warning module is used to integrate multi-source indicators to complete risk assessment.

[0149] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the method described thereon.

[0150] Beneficial Effects: This invention, through model-adaptive feature decoupling and recombination, focuses on characterizing the random perturbations and multimodal uncertainties of future power, while enabling the latent space-time model to focus on characterizing the continuous evolution of potential states driven by trends. Path fusion achieves synergistic unification of the two types of modeling results, thereby improving the accuracy, robustness, and interpretability of early warning results. Through uncertainty quantification and multidimensional concept drift detection, the system can accurately identify operating condition switching. Combined with feature contribution analysis and drift trigger identification modules, it provides interpretable risk warnings, significantly improving the accuracy and robustness of early warnings. Attached Figure Description

[0151] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0152] Figure 1 This is a schematic diagram of the overall structure of a photovoltaic power early warning method based on a diffusion model and latent space provided by the present invention.

[0153] Figure 2 A schematic diagram of the modeling process for the diffusion probability model provided by this invention.

[0154] Figure 3 This is a schematic diagram of the process for constructing the implicit space-time model provided by the present invention.

[0155] Figure 4 This is a schematic diagram illustrating the feature contribution analysis and drift cause identification provided by the present invention.

[0156] Figure 5 This is a schematic diagram of the risk scoring process provided by the present invention. Detailed Implementation

[0157] This embodiment provides a photovoltaic power early warning method based on a diffusion model and latent space, such as... Figure 1 As shown, it includes:

[0158] Step 1: Multi-source data acquisition, alignment and standardized modeling. Information is obtained from multi-source data related to photovoltaic power generation, and time adjustment, missing value repair and anomaly removal are performed on the multi-source data.

[0159] Step 2, Future Power Probability Scenario Generation: Based on historical data, future power trends are generated. Through forward noise addition and reverse noise reduction, future uncertainties and multimodal characteristics are captured, providing basic data samples for further uncertainties.

[0160] Step 3: Based on the multi-scenario samples and historical data generated by the diffusion model, a latent space-time model is constructed. An encoder maps historical data to the latent state space, describing the dynamic stochastic relationship between environmental drivers and power generation. A decoder then continuously generates future latent states and power trends.

[0161] Step 4: Based on the future power trend, further integrate the latent state trajectory and corresponding power trend output by the latent space-time model, and perform joint statistical analysis on the output data of the two models. Merge the power path decoded by the latent space-time model with the power path directly generated by the diffusion model. Calculate statistical quantities such as sample mean, variance, quantiles, and confidence intervals to quantify the fluctuation range and uncertainty of future values.

[0162] Step 5, Concept Drift Detection: By detecting the changing trend of the data, the triggering factor is located, and further change point detection technology is used to output the drift trigger period and its corresponding time position.

[0163] Step 6, Feature Contribution Analysis and Drift Cause Identification. Key causes and drift intensity are identified by detecting changes in feature contribution.

[0164] Step 7: Risk assessment and tiered early warning. Calculate a comprehensive risk score by combining uncertainty indicators and final drift intensity.

[0165] Step 1 includes:

[0166] Step 1.1: Collect data from multiple sources;

[0167] Collect multi-source time-series data related to photovoltaic power prediction and early warning. This multi-source time-series data includes at least meteorological data, electrical operation data, and operation and maintenance status data. For example... Figure 1The meteorological data includes irradiance, ambient temperature, component temperature, wind speed, cloud cover, humidity, and air pressure; the electrical operation data includes voltage, current, active power, reactive power, and inverter operating status; the operation and maintenance status data includes start / stop status, protection actions, shielding status, cleaning status, power limiting status, and control commands.

[0168] Set time The multi-source observation data are as follows:

[0169] ;

[0170] in, Represents an environment variable vector. Represents a vector of power and electrical quantities. This represents the device and its operational status vector.

[0171] Step 1.2, Alignment and Standardization;

[0172] like Figure 1 As shown, unified time alignment, missing value repair, outlier removal, and normalization are performed on multi-source data. The standardized input representation is as follows:

[0173] ;

[0174] in, and These represent the mean and standard deviation obtained from the training samples, respectively.

[0175] Based on this, a sliding time window is used to construct the historical input sequence:

[0176] ;

[0177] The corresponding future target prediction sequence is represented as follows:

[0178] ;

[0179] in, The historical window length is designed to have a value range of [12, 168]. To predict the step size for the future, Indicated by time The future as a starting point Step photovoltaic power output sequence. Indicates the predicted time Photovoltaic power output value at the location, ;therefore, Indicates from the current moment Towards the Future The predicted photovoltaic power at each time step.

[0180] This step outputs a standardized historical input sequence. and future target sequence and the historical input sequence Step 1.3 involves decoupling and recombining model-adaptive features.

[0181] Step 1.3, Decoupling and Recombination of Model-Adaptive Features;

[0182] Obtaining historical input sequences Subsequently, in order to adapt the multi-source input features to the modeling requirements of random perturbations and multimodal distributions, as well as the modeling requirements of the latent space-time model for continuous evolution states, model-adaptive feature decoupling and recombination processing is performed on each input feature in the historical input sequence.

[0183] Let the historical input sequence be... The Middle The characteristic variables are First, calculate the volatility index of this feature within the historical window. :

[0184] ;

[0185] in, Indicates the first The average value of each feature variable within a historical window.

[0186] Calculate the first Time change indicators of each characteristic variable :

[0187] ;

[0188] in, It is used to characterize the magnitude of change of this feature between adjacent time points.

[0189] Calculate the first Time-related indicators between characteristic variables and power output :

[0190] ;

[0191] in, This represents the function for calculating the correlation coefficient. Indicates time Photovoltaic power output.

[0192] Constructing a feature-adaptive discriminant function :

[0193] ;

[0194] in, These are non-negative weighting coefficients. The larger the value, the more random perturbation the feature exhibits; The smaller the value, the more trend-driven the feature is.

[0195] Based on the discrimination threshold The input features are divided into trend-driven feature sets. and random perturbation feature set :

[0196] ;

[0197] in, Used as input to the latent space-time model to characterize the potential state evolution process of photovoltaic power; This is used as a conditional input to characterize the random fluctuations and multimodal uncertainties of future power.

[0198] Equipment operating status, limited-emission status, protection actions, and control commands are not included in the above two types of continuous feature classification, but are instead coded separately as operating status codes. , as a shared constraint input with the latent space-time model.

[0199] Simultaneously, discrete state variables such as equipment operating status, limited-emission status, protection actions, and control commands are constructed into operating status codes. And generate a model-adapted input representation:

[0200] ;

[0201] in, The model adapts the input representation, which serves as the joint input to the subsequent latent space-time model.

[0202] Through the above feature decoupling and recombination process, trend-driven features are mainly used for latent state modeling, random perturbation features are mainly used for probabilistic scenario generation, and running state features are used as shared constraint information to participate in the modeling process of the two models. This improves the collaborative modeling capability with the latent space-time model and avoids the problem of insufficient model adaptability caused by direct mixing of different types of features as input.

[0203] Step 2 includes:

[0204] Step 2.1: Generate future power probability scenarios;

[0205] At any moment Model adaptation input representation In the selection of random perturbation feature set and running status coding As a conditional input, learn the future Photovoltaic power output sequence Conditional probability distribution:

[0206] ;

[0207] in, For parameters, This represents a conditional probability distribution as defined.

[0208] To characterize the multimodal features of future power output, such as Figure 2 As shown, this invention employs a diffusion probability modeling approach, learning the future power distribution through progressive noise addition and denoising. The forward diffusion process is represented as:

[0209] ;

[0210] in, Represented as in the known number Step state Under the conditions, the first Step state The conditional probability distribution; For the first Step-diffusion noise figure, It is the identity matrix. Indicates a Gaussian normal distribution;

[0211] Its equivalent reparameterization is expressed as:

[0212] ;

[0213] in, Indicates the sequence from the initial sequence to the nth. The proportion of signal components retained after step diffusion. Indicates random noise; This can be represented as raw data (initial state) or a target sequence without noise.

[0214] The inverse denoising process recovers the target distribution under given historical conditions, and its training objective is:

[0215] ;

[0216] in, This represents a denoising network. This represents the mathematical expectation operator; this loss function is typically used to constrain the prediction error of the denoising network on noise or target data, enabling the model to progressively learn the mapping relationship from noise to the true data distribution.

[0217] To express the temporal dependencies between future multi-step outputs, the joint conditional distribution of future power is recursively expanded over the prediction step:

[0218] ;

[0219] in, Represented as from time At the time A sequence of photovoltaic power over a continuous time period;

[0220] Multiple sets of future power generation scenarios can be obtained through repeated sampling:

[0221] ;

[0222] in, Indicates the first A future power path, The value can be 1 to N. This is represented as the number of sample paths;

[0223] The set of scenarios output by this step represents the various possible evolutionary outcomes of future power under the same historical conditions, providing an external stochastic reference for the state evolution modeling of the subsequent latent space-time model, and providing a sample basis for uncertainty quantification.

[0224] The learning rate for the diffusion model can take values ​​ranging from 1 to 10. It can obtain high-quality probabilistic scene generation results with reasonable computational cost.

[0225] Step 3 includes:

[0226] Step 3.1, construct the latent space-time model;

[0227] Constructing a latent space-time model based on model-adapted input representation, such as Figure 3 As shown, this invention further establishes a latent space-time model to describe the dynamic relationship between environmental driving forces and power evolution at the latent state level. The generated multi-scenario probabilistic paths are used to characterize the random perturbation boundaries of future power and can serve as consistency reference information for latent state transitions or decoding outputs of the latent space-time model.

[0228] First, the trend-driven feature set is encoded using an encoder. and running status coding Mapping to the latent state space:

[0229] ;

[0230] in, For a moment The potential state representation, For encoder functions, These are encoder parameters.

[0231] The potential state transitions are modeled in a probabilistic form to characterize the stochasticity of the future process:

[0232] ;

[0233] in, This indicates the central tendency of potential state transitions. This indicates the uncertainty of the potential state at that moment.

[0234] The customizations are as follows:

[0235] ,

[0236] ;

[0237] in, Represents the mean mapping function, This represents the covariance mapping function, and the corresponding network layer number is controlled within [1,3]. This represents the external driving vector, which encompasses meteorological environmental variables, power grid dispatching instructions, and equipment control signals.

[0238] The decoder maps the latent states back to the power output space at time 1000. The predicted photovoltaic power is:

[0239] ;

[0240] in, This represents the decoder function.

[0241] This step outputs the future potential state trajectory and the corresponding power path, which, together with the generated future power probability path, serve as the input to the uncertainty quantification module in step 4.

[0242] Among them, the range of the model training learning rate is within This helps avoid training instability and gradient explosion.

[0243] Step 4 includes: jointly fusing the generated future power probability path with the future power path output by the latent space-time model to form a joint power path sample set;

[0244] Step 4.1, Multi-scenario Statistics;

[0245] like Figure 1 As shown, assuming that at a given time, the path generated and the output path of the latent space time model are obtained together. Given a sample of future power paths, with a corresponding value range of [100, 5000], the sample mean at any prediction time is:

[0246] ;

[0247] in, Indicates the first The future power path at the prediction time The power value at that location.

[0248] Step 4.2, uncertainty quantification;

[0249] At any moment Place, The variance of the power path samples is:

[0250] ;

[0251] in, Used to quantify the magnitude of uncertainty in the prediction result at that moment.

[0252] At the predicted time At the point, the empirical distribution of the first quantiles are defined as:

[0253] ;

[0254] in, This represents the quantile probability level, with a range controlled within [0.01, 0.20]. Indicates by The empirical distribution function constructed from the sample paths, Represented as a power value variable; This is represented as the infimum operation;

[0255] Therefore, a confidence interval is constructed:

[0256] ;

[0257] in, Represented as time The confidence interval, Represented as time The 5th percentile (lower quantile); Represented as time The 95th percentile (upper quantile value); Represented as time The width of the interval (the width of uncertainty);

[0258] The interval width is defined as:

[0259] ;

[0260] The wider the interval, the more uncertain the future power.

[0261] This step outputs uncertainty indicators such as sample mean, variance, quantile interval, and interval width, which serve as inputs for concept drift detection in step 5 and risk assessment in step 7.

[0262] Step 5 includes:

[0263] Step 5.1, Concept drift detection;

[0264] like Figure 1 As shown, to identify changes in model applicability conditions and shifts in system operating mechanisms, this invention performs concept drift identification based on prediction residuals, distribution shifts, and change point detection. The prediction residual is defined as:

[0265] ;

[0266] in, For a moment The predicted residuals Indicates time The actual observed power value, This represents the time obtained from the joint output with the latent space-time model. The predicted power value.

[0267] When the mean, variance, or local statistical behavior of the residuals consistently deviates from the reference level, the system can be considered to have experienced a drift trend. If the current window deviates significantly in both residual statistics and distribution shift, it is considered a high-confidence period for conceptual drift. To enhance the ability to locate trigger points, change point detection methods can be combined to identify cumulative changes in residuals and turning points in local means.

[0268] Step 6 includes:

[0269] Step 6.1, Feature contribution analysis;

[0270] like Figure 4 As shown, after detecting concept drift, the contribution of trend-driven features, random disturbance features, and operational status features to changes in prediction bias or uncertainty is further analyzed to pinpoint the key causes of drift. Let the set of variable nodes be: ;in Indicates the first The characteristic variables include irradiance, temperature, cloud cover, wind speed, voltage, current, inverter status, protection actions, power limiting commands, and power output.

[0271] The reference window is selected as the stable operating window before the drift trigger (such as the past 24 hours), and the current window is the recent window containing the trigger time (such as the past 1 hour).

[0272] Based on the current drift window For samples within the range, each feature is calculated using model interpretation methods (e.g., SHAP value, feature gradient contribution, or local perturbation analysis). Contribution to model prediction residuals or uncertainty indicators The change in feature contribution relative to a reference window (such as the stable operating window before drift) is defined as:

[0273] ;

[0274] in This represents the average contribution of the same feature within the reference window. Sort by largest to smallest, and select the top... These features serve as a set of key inducing variables. .

[0275] Step 6.2, Drift Cause Analysis and Intensity Quantification; further define the first... The drift intensity index for each window is:

[0276] ;

[0277] in Represents the set of key inducing variables. This indicates the degree of deviation of key factors from the reference operating condition (e.g., the standardized deviation of eigenvalues). and These are the weighting coefficients. The first term measures the overall magnitude of change in the feature contribution, and the second term measures the degree of anomalousness in the values ​​of the key factors themselves.

[0278] Based on the characteristic category of the key inducing variables, drift inducing factors can be divided into trend evolution inducing factors, random disturbance inducing factors, and operational state constraint inducing factors.

[0279] In practical implementation, feature contribution can be calculated using model-independent interpretation methods (such as SHAP and LIME) or model-dependent gradient or attention mechanisms. For example, under normal clear-sky conditions, irradiance typically contributes the most; under cloud shadow disturbance conditions, the contributions of cloud cover and irradiance increase significantly; under equipment thermal derating conditions, the contribution of component temperature or inverter temperature rise increases; and under control constraint conditions, the contribution of emission limitation commands becomes dominant. By calculating the feature contribution distribution of each window through a sliding window, it is possible to observe which features show significant changes in contribution before and after drift. The identification of key inducing variables is based on the ranking of contribution changes; the larger the absolute value of the change, the stronger the influence of the feature on the current drift. This step outputs key inducing variables, a description of feature contribution changes, and a drift intensity index, which serve as inputs to the risk assessment and graded early warning module.

[0280] Step 7 includes:

[0281] Step 7.1, Risk Assessment;

[0282] After obtaining the uncertainty index and drift intensity, such as Figure 5 As shown, this invention performs a joint assessment of multi-source risk information to form a unified risk score. The risk score comprehensively considers the volatility of future power prediction distribution, quantile interval width, drift intensity, real-time residual amplitude, and the degree of anomaly of key factors, and can be expressed as:

[0283] ;

[0284] in, For a moment Comprehensive risk score; For a moment Prediction standard deviation (characterizing the degree of dispersion); For a moment Confidence interval width (characterizing the range of uncertainty); For a moment Drift intensity index (reflects changes in feature contribution and the degree of deviation of key factors); Indicates time Predicting residuals; Indicates time The intensity of key anomalies (weighted by the change in contribution or deviation of eigenvalues ​​of key inducing variables), wherein the intensity of key anomalies can be obtained by weighting the change in contribution and deviation of eigenvalues ​​of the key inducing variables identified in step 6. These are non-negative weighting coefficients used to adjust the contribution of each risk component to the overall score.

[0285] Step 7.2, tiered early warning;

[0286] like Figure 5 As shown, based on different power plants, seasons, and operating conditions, the weighting coefficients and scoring thresholds can be calibrated according to historical statistical results or operational experience, and risks can be classified into low risk, moderate risk, relatively high risk, and high risk levels. During the early warning output phase, this invention does not only output a single level label, but also simultaneously outputs the potential future power fluctuation range, risk level, main inducing variables, description of changes in characteristic contributions, and corresponding handling suggestions. Differentiated response strategies can be generated for different inducing factor types. Specific Implementation

[0287] 1. Data Acquisition and Preprocessing:

[0288] This embodiment uses historical operating data from a centralized photovoltaic power station as an example to illustrate the method of the present invention. The data used covers the entire operating cycle of 2023, including different seasons such as spring, summer, autumn, and winter, as well as various typical operating conditions such as clear skies, thin clouds, thick clouds, partial shading, power limiting control, and equipment status changes. It can comprehensively reflect the dynamic changes in photovoltaic power under complex environments. The types of data collected include: meteorological data (irradiance, ambient temperature, module temperature, wind speed, cloud cover, humidity, and air pressure), electrical operation data (voltage, current, active power, reactive power, and inverter operating status), and operation and maintenance status data (start-stop status, protection actions, cleaning status, shading status, power limiting status, and dispatch control commands).

[0289] For the collected multi-source time-series data, the training set, validation set and test set are divided according to the time order, with the training set accounting for 70%, the validation set accounting for 10% and the test set accounting for 20%, in order to avoid future information leakage caused by random division.

[0290] The raw data is then preprocessed, including time alignment, missing value repair, outlier removal, normalization, and sliding window construction.

[0291] 2. Experimental Results and Analysis:

[0292] 2.1 Evaluation Indicators

[0293] To verify the effectiveness of the method of the present invention, the following indicators are used to evaluate the model in this embodiment: Anomaly Detection Accuracy, Feature Extraction Sensitivity, Risk Assessment Composite Score, and Model Stability Index.

[0294] 2.2 Comparative Experiment Design

[0295] Set up the following comparison models:

[0296] (1) Traditional LSTM prediction model: It only uses historical time series data to predict future power and combines a fixed threshold to make early warning judgment.

[0297] (2) GRU prediction model: The historical power and environmental data are modeled using a gated cyclic cell network, and then the early warning output is given based on the prediction deviation.

[0298] (3) CNN-LSTM hybrid model: First, use the convolutional structure to extract local features, and then combine LSTM to complete the time series prediction and early warning judgment.

[0299] (4) Rule threshold + statistical feature model: anomaly detection and early warning judgment based on power residual, volatility and key environmental variable thresholds.

[0300] (5) The model of this invention adopts a closed-loop early warning framework that includes conditional diffusion probability scenario generation, latent space temporal world model, uncertainty quantification, concept drift detection, temporal causal graph causal analysis, and risk scoring and adaptive updating.

[0301] All models are trained and tested on the same dataset.

[0302] 2.3 Results Analysis

[0303] 2.3.1 Performance comparison of different models in anomaly identification and power prediction tasks, as shown in Table 1.

[0304] Table 1

[0305]

[0306] The method of this invention outperforms traditional comparative methods in terms of anomaly detection accuracy and feature extraction sensitivity, indicating that the invention can more effectively identify changes in complex operating conditions and more sensitively capture key features such as irradiance, temperature, emission limitation status, and equipment status changes.

[0307] This invention provides a photovoltaic power early warning method and system based on a diffusion model and hidden space. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A photovoltaic power early warning method based on diffusion model and latent space, characterized in that, Includes the following steps: Step 1: Obtain meteorological data, electrical operation data, and operation and maintenance status data from multi-source data related to photovoltaic power generation. Perform time alignment, missing value repair, anomaly removal, and standardization on the multi-source data. Based on feature volatility, time variability, and power correlation, construct a feature adaptation discriminant function to decouple multi-source features into trend-driven features, random disturbance features, and operation status features, forming a model adaptation input representation. Step 2: Select random perturbation features and operating state features as conditional inputs, and use the forward noise addition and reverse noise reduction process to generate a multi-scenario probability path for future power in order to capture the uncertainty and multimodal features of future power. Step 3: Select trend-driven features and operating state features as inputs to the latent space-time model, and combine the output multi-scenario samples as uncertainty references. The encoder maps the samples to the latent state space to establish latent state transition relationships, and the decoder generates future latent state trajectories and corresponding power trends. Step 4: Merge the generated future power probability path with the power path decoded by the latent space-time model to form a joint power path sample set; perform statistical analysis on the joint sample set to calculate the sample mean, variance, quantiles, confidence intervals and interval widths, and quantify the range of future power fluctuations and prediction uncertainty. Step 5, Concept Drift Detection: By detecting the changing trend of the data, the triggering factor is located, and then the change point detection technology is further applied to output the drift triggering period and the corresponding time position. Step 6, Feature contribution analysis and drift cause identification: Locate key causes and drift intensity by detecting changes in feature contribution; Step 7, Risk Assessment and Tiered Early Warning: Calculate a comprehensive risk score by combining uncertainty indicators and final drift intensity.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Collect data from multiple sources; Collect multi-source time-series data related to photovoltaic power prediction and early warning, including meteorological data, electrical operation data, and operation and maintenance status data; The meteorological data includes irradiance, ambient temperature, component temperature, wind speed, cloud cover, humidity, and air pressure. The electrical operating data includes voltage, current, active power, reactive power, and inverter operating status parameters. The operation and maintenance status data includes start / stop status, protection action, blocking status, cleaning status, limited generation status, and control commands; Let multi-source observation data at time t be used. for: ； in, This represents the vector of environmental variables at time t. This represents the power and electrical quantity vector at time t. This represents the device and its operational status vector at time t; Step 1.2, Alignment and Standardization; Perform unified time alignment, missing value repair, outlier removal, and normalization on multi-source data. The input at time t is then standardized. Represented as: ； in, and These represent the mean and standard deviation obtained from the training samples, respectively. Constructing the historical input sequence at time t using a sliding time window : ； The corresponding future target prediction sequence is represented as follows: ； in, It is the length of the historical window. It is the future predicted step size. Indicated by time The future as a starting point Step photovoltaic power output sequence; Indicates the predicted time Photovoltaic power output value at the location, ; This step outputs a standardized historical input sequence. and future target sequence and the historical input sequence Step 1.3 involves decoupling and recombining model-adaptive features. Step 1.3, Decoupling and Recombination of Model-Adaptive Features; Let the historical input sequence be... The Middle The characteristic variables are First, calculate the volatility index of the feature within the historical window. : ； in, Indicates the first The average value of each feature variable within a historical window; Calculate the first Time change indicators of each characteristic variable : ； Calculate the first Time-related indicators between characteristic variables and power output : ； in, This represents the function for calculating the correlation coefficient. Indicates time Photovoltaic power output; Constructing a feature-adaptive discriminant function : ； in, These are non-negative weighting coefficients; Based on the discrimination threshold The input features are divided into trend-driven feature sets. and random perturbation feature set : ； Simultaneously, discrete state variables are constructed as running state codes. And generate a model-adapted input representation: ； in, The model adapts to the input representation.

3. The method according to claim 2, characterized in that, Step 2 includes the following steps: Step 2.1: Generate future power probability scenarios; At any moment Model adaptation input representation In the selection of random perturbation feature set and running status coding As a conditional input, learn the future Photovoltaic power output sequence Conditional probability distribution: ； in, For parameters, Represents the conditional probability distribution defined; Using a diffusion probability modeling approach, the future power distribution is learned through progressive noise addition and denoising. The forward diffusion process is represented as follows: ； in, Represented as in the known number Step state Under the conditions, the first Step state The conditional probability distribution; For the first Step-diffusion noise figure, It is the identity matrix. Indicates a Gaussian normal distribution; Equivalent reparameterization is represented as: ； in, Indicates the sequence from the initial sequence to the nth. The proportion of signal components retained after step diffusion. Indicates random noise; Represented as raw data or a target sequence without noise; The inverse denoising process restores the target distribution under given historical conditions; the loss function of the diffusion model. Represented as: ； in, This represents a denoising network. Represents the mathematical expectation operator; The joint conditional distribution of future power is recursively expanded according to the prediction step: ； in, Represented as from time At the time A sequence of photovoltaic power over a continuous time period; Two or more future power generation scenarios are obtained through repeated sampling, at any given time. The generated future power path sample set for: ； in, Indicates the first A future power path, The value ranges from 1 to N, where N represents the total number of sample paths.

4. The method according to claim 3, characterized in that, Step 3 includes the following steps: Step 3.1, construct the latent space-time model; First, the trend-driven feature set is encoded using an encoder. and running status coding Mapping to the latent state space: ； in, For a moment The potential state representation, For encoder functions, For encoder parameters; The potential state transition is modeled using a probabilistic form: ； in, This indicates the central tendency of potential state transitions. This represents the uncertainty of the potential state at that moment; it is defined as: ， ； in, Represents the mean mapping function, Represents the covariance mapping function. Represents the external driving vector; The decoder maps the latent states back to the power output space at time 1000. Photovoltaic power forecast for: ； in, This represents the decoder function.

5. The method according to claim 4, characterized in that, Step 4 includes the following steps: Step 4.1, Multi-scenario Statistics; Suppose that at a given time, the path generated and the output path of the latent space time model are obtained together. Given a sample of future power paths, the sample mean at any prediction time is... for: ； in, Indicates the first The future power path at the prediction time Power value at; Step 4.2, uncertainty quantification; At any moment Place, Variance of power path samples for: ； At the predicted time At the point, the empirical distribution of the first quantiles Defined as: ； in, Indicates the quantile probability level. Indicates by The empirical distribution function constructed from the sample paths, Represented as a power value variable; Indicates the infimum operation; Construct confidence intervals: ； in, Represented as time The confidence interval, Represented as time 5% quantile; Represented as time 95th percentile; Represented as time The interval width; The interval width is defined as: 。 6. The method according to claim 5, characterized in that, Step 5 includes the following steps: Step 5.1, Concept drift detection; The prediction residual is defined as: ； in, For a moment The predicted residuals Indicates time The actual observed power value, This represents the time obtained from the joint output with the latent space-time model. The predicted power value.

7. The method according to claim 6, characterized in that, Step 6 includes the following steps: Step 6.1, Feature contribution analysis; After detecting concept drift, we further analyze the contribution of trend-driven features, random disturbance features, and operational status features to changes in prediction bias or uncertainty, thereby identifying the key causes of drift. Let the set of variable nodes V be: ; in Indicates the first There are 1 feature variables; n represents the total number of feature variables; Based on the current drift window For the samples within the dataset, each feature is calculated using a model interpretation method. Contribution to model prediction residuals or uncertainty indicators Define the change in feature contribution relative to the reference window. for: ； in The average contribution of the same feature under the reference window; Sort by largest to smallest, and select the top... These features serve as a set of key inducing variables. ; Step 6.2, Drift Cause Analysis and Intensity Quantification; Definition of the first Drift intensity index of each window for: ； in, This indicates the degree of deviation of key factors from the reference operating condition. and These are the weighting coefficients.

8. The method according to claim 7, characterized in that, Step 7 includes the following steps: Step 7.1, Risk Assessment; After obtaining uncertainty indicators and drift intensity, multi-source risk information is jointly assessed to form a unified risk score; The risk score comprehensively considers the volatility of the future power prediction distribution, the width of the quantile interval, the drift intensity, the real-time residual amplitude, and the degree of anomaly of key factors, and is expressed as follows: ； in, For a moment Comprehensive risk score; Indicates time Forecast standard deviation; Indicates time Confidence interval width; Indicates time Drift intensity index; Indicates time Intensity of key anomalous factors; These are non-negative weighting coefficients; Step 7.2, tiered early warning; Based on different power plants, seasons, and operating conditions, the weighting coefficients and scoring thresholds are calibrated according to historical statistical results or operating experience, and the risks are classified into low risk, general risk, relatively high risk, and high risk levels.

9. A photovoltaic power early warning system based on a diffusion model and latent space, implemented using the method described in any one of claims 1 to 7, characterized in that, include: The multi-source data acquisition, alignment, and standardized modeling module is used for cleaning and standardizing multi-source data and initiating a historical project input window. The model-adaptive feature decoupling and recombination module is used to divide historical input features into trend-driven features, random disturbance features, and operating state features based on feature volatility, time variability, and power correlation, and generate model-adaptive input representations. The future power probability scenario generation module is used to generate future power multimodal probability scenario samples based on random perturbation features and operating state encoding. The latent space temporal model building module is used to map trend-driven features and running state encodings to the latent state space. The dual-model path fusion and uncertainty quantification module is used to perform joint statistical analysis on multi-path samples output by the implicit space-time model to quantify the range of uncertainty in future predictions. The concept drift detection module is used to identify concept triggers based on the joint analysis of residuals and distribution shifts. The feature contribution analysis and drift cause identification module is used to detect and locate key causes and drift intensity through changes in feature contribution. The risk assessment and tiered early warning module is used to integrate multi-source indicators to complete risk assessment.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the method as described in any one of claims 1 to 8.

Images