Solar irradiance diffuse scene prediction model and post-training method thereof

Abstract

The application discloses a solar irradiance diffusion scene prediction model and a post-training method thereof. The post-training method comprises the following steps: taking a condition diffusion model trained as a basic probability scene generator to generate a plurality of future scenes; calculating a plurality of decision utility groups: fusing basic distribution quality, identifying climbing events, controlling under-prediction risks, key period reliability and scheduling cost design decision utility functions; taking the plurality of candidate scenes generated under the same input condition as a whole in a group, and constructing a list type target distribution according to the relative decision utility of the candidate scenes; and performing list type decision alignment diffusion post-training: introducing reference model constraints, physical consistency constraints and probability calibration constraints in addition to decision alignment loss. The solar irradiance diffusion scene prediction model is upgraded from simple statistical fitting to decision utility alignment, so that the model output is more in line with the needs of power system scheduling and energy storage control.

Description

Technical Field

[0001] This invention relates to the field of solar irradiance or photovoltaic power prediction technology, and more specifically, to a solar irradiance diffusion scenario prediction model and its subsequent training method. Background Technology

[0002] With the high proportion of new energy sources connected to the grid, solar irradiance or photovoltaic power prediction has become a key technology in power dispatch, energy storage control, reserve capacity configuration, and power market operation. Existing methods for solar irradiance or photovoltaic power prediction mainly include point prediction methods, probabilistic prediction methods, scenario generation methods for power system dispatch, and post-processing methods for prediction based on cost or risk. Among them, point prediction methods typically output a single path at one or more future times; probabilistic prediction methods typically output prediction intervals, quantiles, probability distributions, ensemble predictions, or multiple future scenarios; scenario generation methods typically input generated wind power, photovoltaic, and load scenarios into stochastic optimization, robust optimization, chance-constrained optimization, energy storage dispatch models, or power market bidding models to generate downstream dispatch or trading decisions.

[0003] Among the methods described above, the closest prior art to this invention is a method for probabilistic prediction and scenario generation of solar irradiance or photovoltaic power for power system decision-making. This type of method typically first constructs a probabilistic prediction model for solar irradiance or photovoltaic power, outputting prediction intervals, quantiles, probability distributions, ensemble predictions, or multiple future scenarios (see van der Meer DW, Widén J, Munkhammar J. Review on probabilistic forecasting of photovoltaic power production and electricity consumption. Renewable and SustainableEnergy Reviews, 2018, 81: 1484-1512); then, the prediction results are input into stochastic optimization, robust optimization, chance-constrained optimization, energy storage dispatching models, or electricity market bidding models to generate downstream dispatching or trading decisions. Probabilistic prediction models for solar irradiance or photovoltaic power can be implemented using various techniques such as quantile regression, ensemble prediction, Bayesian models, hybrid density networks, normalized flow, VAE, GAN, and diffusion. Among these, the diffusion model generates multiple future scenarios through conditional backsampling, making it particularly suitable for predicting solar irradiance scenarios in high-dimensional regions.

[0004] Existing methods for probabilistic prediction of solar irradiance or photovoltaic power and for generating scenarios still have at least the following problems when facing downstream decision-making tasks such as power system dispatch, energy storage control, reserve capacity configuration, or electricity market bidding:

[0005] 1. The predicted training objectives and the scheduling decision objectives are not completely consistent.

[0006] Existing probabilistic prediction models for solar irradiance or photovoltaic power typically employ statistical error or probabilistic scoring rules such as root mean square error, mean absolute error, negative log-likelihood, continuous sorting probability score, energy score, variogram score, interval coverage, and reliability test for evaluation (see Gneiting T, Raftery A E. Strictly ProperScoring Rules, Prediction, and Estimation. Journal of the American Statistical Association, 2007, 102(477): 359-378; Scheuerer M, Hamill T M. Variogram-Based Proper Scoring Rules for Probabilistic Forecasts of Multivariate Quantities. Monthly Weather Review, 2015, 143(4): 1321-1334). These rules can evaluate prediction accuracy, probabilistic calibration, or scenario distribution quality. However, actual power system dispatching focuses more on reliability during critical periods, identification of ramp-up events, under-prediction risk, reserve capacity requirements, energy storage operating costs, or market bidding revenue. Therefore, more accurate predictions in a statistical sense do not necessarily lead to better downstream scheduling decisions.

[0007] 2. Downstream decision-making value is usually not directly fed back to the scene generation model itself.

[0008] In existing forecast-then-optimize methods, the probabilistic prediction results of solar irradiance or photovoltaic power are usually used as input to an external dispatch optimization model, which then generates a dispatch plan through stochastic optimization, robust optimization, chance-constrained optimization, energy storage dispatch models, or market bidding models. While these methods can utilize probabilistic prediction results for decision-making, the solar irradiance or photovoltaic power probabilistic prediction models themselves are typically still trained according to statistical loss, probability scoring rules, or distribution fitting to the target. Downstream dispatch utility, reserve capacity cost, energy storage operating cost, or market revenue are usually only used for external optimization or effect evaluation after the prediction is completed. No direct post-training feedback mechanism for the sampling probability of the scenario generation model is formed (see Li R, Zhang H, Sun M, Teng F, Wan C, Pineda S, Kariniotakis G. Decision-Oriented Learning for Future Power System Decision-Making underUncertainty. arXiv:2401.03680, 2024).

[0009] 3. Directly introducing cost-sensitive objectives or decision-making rewards may lead to a degradation of probabilistic quality.

[0010] Existing probabilistic prediction models for solar irradiance or photovoltaic power are simply trained by incorporating cost-sensitive losses, scheduling rewards, or reinforcement learning rewards. However, they lack reference model constraints, probabilistic calibration constraints, and physical consistency constraints. This can lead to models that excessively favor specific high-reward outputs, resulting in shifted prediction distributions, distorted confidence intervals, degraded probabilistic calibration, or physically inappropriate outputs. For example, the model might systematically overestimate daytime power output to reduce under-prediction penalties, or produce physically inconsistent scenarios at night, when solar altitude angles are low, or when cloud formations are rapidly changing.

[0011] 4. High-dimensional scene prediction lacks a stable intra-group relative optimization mechanism.

[0012] Regional solar irradiance fields or photovoltaic power probabilistic prediction tasks typically have high-dimensional spatiotemporal structures, and the absolute utility scores of individual candidate scenarios are easily affected by sudden weather changes, local cloud movement, observational noise, or scheduling agent model errors. Existing methods focus more on the statistical error of individual prediction results, the overall score of the scenario set, or inputting the prediction results into an external optimization model.

[0013] 5. The deep probabilistic generative model of the diffusion model has not yet fully incorporated the training mechanism after scheduling utility.

[0014] Existing diffusion model-based probabilistic prediction methods for solar irradiance or photovoltaic power mainly utilize diffusion models for high-resolution probabilistic prediction or multi-sample generation. Their training objectives typically revolve around denoising and reconstruction, score matching, likelihood lower bounds, or statistical distribution fitting. For regional solar irradiance or photovoltaic power probabilistic prediction tasks geared towards scheduling decisions, existing diffusion model methods have not yet adequately provided an independent, stable, and reusable post-training stage that allows the diffusion model's sampling distribution to be simultaneously aligned with scheduling utility, ramp-up event identification, underprediction risk control, critical period reliability, physical consistency, and probabilistic calibration. Summary of the Invention

[0015] To address the problems existing in the prior art, the present invention aims to provide a solar irradiance diffusion scenario prediction model, its subsequent training method, and related computer program products, in order to solve the problem of inconsistency between the statistical fitting target and the actual deployment decision target in existing solar irradiance or photovoltaic power probability prediction models.

[0016] To achieve the above objectives, in a first aspect, the present invention provides a post-training method for a solar irradiance diffusion scenario prediction model, comprising the following steps:

[0017] Multiple future scenes are generated using the conditional diffusion model that has been trained on the basis as the basic probabilistic scene generator; after the conditional diffusion model is trained on the basis, the basic model parameters are frozen and kept from being updated in the later training stage.

[0018] Calculating multi-component decision utility: The decision utility function is composed of a weighted average of multiple utility terms.

[0019]

[0020] in, The mass term of the basic distribution; For the hill climbing event item; This is an asymmetric underpredictable risk item; For reliability items during critical periods; This is the scheduling cost item; The weights of each utility item;

[0021] Multiple candidate scenarios generated under the same input conditions are treated as a whole within a group, and the relative target distribution within the group is constructed based on the relative decision utility of the candidate scenarios.

[0022] Post-training with listwise decision alignment diffusion: In addition to the decision alignment loss, reference model constraints, physical consistency constraints, and probabilistic calibration constraints are introduced. The total loss function after post-training is:

[0023]

[0024] Where: L is the total post-training loss; L align For list-based decision alignment loss; L ref For reference model consistency constraints; L phys For physical consistency constraints; L cal For probability calibration constraints; These are the weight parameters for each constraint term.

[0025] The above technical solution uses a conditional diffusion model as the basic probabilistic scenario generator, leveraging its natural sampling capabilities to generate multiple future scenarios. Post-training with decision utility modifies the diffusion model's sampling bias towards high-value scenarios, enabling the model to characterize the uncertainties of future weather and photovoltaic output. A decision utility function is designed, integrating basic distribution quality, ramp-up event identification, under-prediction risk control, critical period reliability, and scheduling costs, to evaluate the actual value of each candidate prediction scenario for power system dispatch, energy storage control, or market bidding. Based on the intra-group relative utility of multiple candidate scenarios under the same input, a list-style decision alignment post-training objective is constructed to improve the model's probability of generating high-utility scenarios. Reference model constraints, physical consistency constraints, and probabilistic calibration constraints prevent distribution drift, physically unreasonable outputs, and probabilistic calibration degradation caused by post-training. Overall, these measures make the output of the post-trained solar irradiance diffusion scenario prediction model more aligned with the needs of power system dispatch and energy storage control, providing higher reference value for actual deployment decisions.

[0026] In some implementations, the conditional diffusion model includes a forward denoising process and a reverse denoising process: First, the denoised future scene, historical observation features, and diffusion step encoding under the current diffusion step are input into a 3D convolutional input module to obtain low-level spatiotemporal features; the low-level spatiotemporal features are further input into a U-Net encoder to obtain multi-scale encoded features and bottleneck features; the bottleneck features are input into a temporal Transformer module to model cross-temporal dependencies within the prediction window; exogenous conditional features, including satellite, NWP, solar geometry, and time calendar, are processed by a conditional encoder to obtain conditional features, which are then fused with bottleneck features or multi-scale encoded features through a cross-attention module; the fused features are input into a U-Net decoder, and spatial resolution is restored stepwise by combining skip connections from the encoder; finally, the conditional diffusion denoising head outputs noise predictions, clean sample predictions, or velocity variable predictions.

[0027] In some implementations, the loss function for the basic training of the conditional diffusion model is: ,

[0028] Among them, L base Based on the total training loss; L diff For diffusion denoising loss; Lenergy The Energy Score loss is used to evaluate the overall distributional difference between the multi-scenario prediction ensemble and the actual future observations; L variogram The VariogramScore loss is used to evaluate the prediction quality of dependencies between different locations in a multivariate or spatial scene; L l1 The point-state L1 loss is used to constrain local numerical errors; w es w vs w l1 These are the weight parameters for the corresponding loss term.

[0029] In some implementations, the basic distribution quality term constrains the overall similarity between candidate scenarios and actual future observations; the ramp-up event term evaluates whether candidate scenarios can identify rapid increases or decreases in solar irradiance or photovoltaic power; the asymmetric underprediction risk term evaluates the different risks of underprediction and overprediction in the power system; the critical period reliability term evaluates the prediction quality during periods of high irradiance, high photovoltaic output, or high system value; and the dispatch cost term evaluates the actual value of candidate prediction scenarios for energy storage dispatch, reserve capacity allocation, or electricity market bidding.

[0030] In some implementations, the basic distribution quality term is the negative of the mean absolute error between the candidate scenario and the actual future observation; the climbing event term is the average of the soft-F1 scores of the uphill and downhill climbs; the asymmetric underprediction risk term is the negative of the asymmetric prediction risk loss, and the asymmetric prediction risk loss... Where: T is the number of time steps within the prediction window, For spatial grid point set or power station set, For time indexing, Weighting for under-prediction penalties; Over-prediction penalty weights; The magnitude is under-predicted; The over-prediction magnitude is defined as follows: the critical period reliability term is the negative of the average prediction error of the candidate scenario during the critical period; the scheduling cost term is the negative of the scheduling regret value, which is the difference between the actual operating cost of the scheduling plan generated based on the candidate prediction scenario under the real future result and the ex-post optimal cost obtained by directly solving based on the real future result under the same scheduling constraints.

[0031] In some implementations, constructing the relative target distribution within a group includes: converting the candidate scene utility scores {r1, r2, …, rk} under the same input X into a target scene distribution using a softmax function with a temperature parameter.

[0032] ,

[0033] Where qk is the probability of the k-th candidate scene in the target distribution; is the utility temperature parameter, used to control the sharpness of the target distribution; K is the number of candidate scenes.

[0034] In some implementations, the list-based decision alignment loss is:

[0035] ,

[0036] in, Losses due to decision alignment; For the target distribution q and the model-induced distribution The KL divergence between them; α is the weighting parameter of the weighted diffusion denoising loss; For the current model, the candidate scene Y (k) The diffusion denoising loss;

[0037] The reference model consistency constraint limits the deviation of the current model from the frozen reference model; the physical consistency constraint ensures that the predicted scenario meets the basic physical laws of solar irradiance or photovoltaic power; and the probabilistic calibration constraint ensures that the uncertainty of the model output matches the actual error.

[0038] Secondly, the present invention provides a solar irradiance diffusion scenario prediction model, which is obtained by training the conditional diffusion model that has been trained using the post-training method described above.

[0039] In some implementations, the solar irradiance diffusion scenario prediction model outputs one or more results from a set of scenarios for any input X, including multiple future solar irradiance scenarios, probability weights for each scenario, weighted mean prediction, prediction intervals or quantiles, and scenarios available for energy storage scheduling, reserve capacity configuration, ramp-up event warning, or electricity market bidding.

[0040] Finally, the present invention provides a computer program product, which, when run on a processor, executes the post-training method for the solar irradiance diffusion scenario prediction model as described above.

[0041] Compared with the prior art, the present invention has the following technical effects:

[0042] (1) This invention improves the solar irradiance diffusion scenario prediction model from simple statistical fitting to decision utility alignment, so that the model output is more in line with the needs of power system dispatch and energy storage control.

[0043] (2) This invention generates multiple future scenarios through a conditional diffusion back sampling process, which makes the prediction results naturally have diversity and conditional probability interpretation capabilities.

[0044] (3) This invention enables the efficient use of diffusion sampling scenarios to obtain a higher generation tendency under the same input by training through relative listwise within the group, thereby improving the prediction value of climbing events, under-predicted risks and critical periods.

[0045] (4) This invention reduces the risks of reward hacking, distribution drift and probability distortion by using reference models, physical consistency and calibration constraints. Attached Figure Description

[0046] The invention, its features and advantages will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0047] Figure 1 This is a flowchart illustrating the training and operation of the solar irradiance diffusion scenario prediction model in an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of the basic conditional diffusion model in an embodiment of the present invention.

[0049] Figure 3 This is a flowchart of the training process after Listwise decision alignment diffusion in an embodiment of the present invention. Detailed Implementation

[0050] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but these are not intended to limit the scope of the invention.

[0051] In this embodiment of the invention, "probabilistic prediction" refers to the uncertainty information of future solar irradiance or photovoltaic power output by the prediction model, including but not limited to prediction intervals, quantiles, probability distributions, ensemble prediction results, or multiple future scenarios. In this embodiment of the invention, "scenario" refers to a possible realization of the solar irradiance field or photovoltaic power sequence of a target area within a future period.

[0052] See Figure 1 This invention provides a method for training and running a computer-based solar irradiance diffusion scenario prediction model. The method includes a basic condition diffusion scenario prediction model training stage, a candidate scenario generation stage, a decision utility evaluation stage, a listwise decision alignment post-training stage, a stability constraint stage, and a prediction result output stage.

[0053] For ease of explanation, the following symbols are defined first:

[0054] X: Model input, representing a set of historical observation data and exogenous features, including at least one of historical solar irradiance, historical photovoltaic power, satellite cloud images, numerical weather prediction data, solar altitude angle, solar azimuth angle, and time calendar features.

[0055] Y gt : Real future observation results, representing the real future solar irradiance field or photovoltaic power sequence corresponding to input X in the training samples.

[0056] Y: Future prediction target, representing the regional solar irradiance field or photovoltaic power sequence over multiple future time steps. In regional scenario prediction, Y can be represented as a three-dimensional tensor:

[0057] ,

[0058] Where T represents the prediction time step, and H and W represent the height and width of the spatial grid.

[0059] Y (k) : The k-th candidate future scene, where k=1,2,…,K, and K represents the number of candidate scenes generated under the same input conditions.

[0060] f θ : The conditional probability prediction model to be trained, where θ represents the model parameters.

[0061] : The base model after basic training is completed, where θ0 represents the parameters of the base model.

[0062] f ref : Freeze the reference model, based on the base model The result is obtained by copying and freezing, which is used to limit the current model from shifting too much relative to the base model during post-training.

[0063] : The conditional generation distribution of the current model for the future scenario Y under the given input X.

[0064] r k The decision utility score for the k-th candidate scenario can also be understood as a reward or a comprehensive evaluation score. In the patent text, it is referred to as the "decision utility value".

[0065] q k : The target scenario distribution induced by the decision utility score represents the target relative probability that the model wants to assign to the k-th candidate scenario.

[0066] The model scene distribution induced by the current model's generated score represents the probability that the current model considers the k-th candidate scene to be relatively more likely to be generated.

[0067] Step S1: Input data acquisition and preprocessing

[0068] S101. Obtain at least one of the following for the target area: historical solar irradiance data, photovoltaic power data, satellite cloud image data, numerical weather prediction (NWP) data, solar geometric data, and time calendar features.

[0069] S102. In one implementation, the Global Horizontal Irradiance (GHI) is converted into a Clear Sky Index:

[0070] ,

[0071] Where: k t GHI represents the clear sky index at time t. t GHI represents the global horizontal irradiance observed at time t. clear,t This represents the theoretical irradiance at the same location and time under clear sky conditions. When GHI clear,t When the value is less than a preset threshold, a nighttime or low solar altitude angle mask is set for that moment to avoid meaningless ratios at nighttime or when the solar altitude angle is too low affecting model training.

[0072] The input data is standardized to a fixed time interval and a fixed spatial grid. In one implementation, the time interval is 15 minutes, the spatial grid is 64×64, the input window is the past 4 hours, and the prediction window is the next 4 hours. The aforementioned time interval, spatial grid size, input window length, and prediction window length can all be adjusted according to the specific application scenario.

[0073] Step S2: Training the basic condition diffusion scenario prediction model

[0074] Constructing a basic condition diffusion scenario prediction model Its input consists of historical observations and exogenous features X, and its output is the conditionally generated distribution of the future solar irradiance field or photovoltaic power trajectory: .

[0075] In a preferred embodiment, the basic conditional diffusion scenario prediction model employs a conditional diffusion model. In low-computing-power or high-real-time deployment scenarios, the conditional diffusion model can be distilled into an explicit multi-scenario prediction head or a few-step consistency model. The conditional diffusion model includes a forward noise addition process and a reverse denoising process. The forward noise addition process is used to progressively perturb the real future scene into noise samples, while the reverse denoising process is used to progressively recover the future scene from the noise given the input condition X.

[0076] The forward noise addition process can be represented as:

[0077] ,

[0078] Where: Yt Y0 represents the noisy sample at the t-th diffusion step; Y0 represents the real future scene, i.e., the real label without noise; t represents the diffusion time step or noise level. Noise sampled from a standard Gaussian distribution; The diffusion noise scheduling parameter controls the ratio of the original signal retained to the added noise at the t-th diffusion step. The inverse denoising process is implemented by a denoising network. This denoising network operates on a given noisy sample Y. t Given a diffusion step t and conditional input X, predict noise, clean samples, or velocity variables. For example, when using the noise prediction form, it can be expressed as:

[0079] ,

[0080] in, This represents the output of a denoising network with parameter θ.

[0081] See Figure 2 The basic conditional diffusion scenario prediction model may include a conditional input encoding module 210, a diffusion step encoding module 220, a diffusion denoising network 230, and a conditional diffusion output head 240. Specifically, the conditional input encoding module 210 encodes at least one of historical solar irradiance, historical photovoltaic power, satellite cloud imagery, NWP data, solar geometric features, and time calendar features to obtain conditional features; the diffusion step encoding module 220 encodes the current diffusion step (t) to obtain diffusion step features; the diffusion denoising network 230 extracts spatiotemporal features and performs inverse denoising given a noisy future scenario, conditional features, and diffusion step features; and the conditional diffusion output head outputs noise predictions, clean sample predictions, or velocity variable predictions.

[0082] In a preferred embodiment, the diffusion denoising network includes a 3D convolutional input module 231, a U-Net encoder 232, a temporal Transformer bottleneck module 233, a cross-attention fusion module 234, and a U-Net decoder 235. The 3D convolutional input module 232 is used to extract short-term cloud motion and local spatiotemporal texture; the U-Net encoder is used to extract multi-scale spatial features; the temporal Transformer bottleneck module 233 is used to model cross-temporal dependencies within the prediction window; the cross-attention fusion module 234 is used to fuse exogenous features such as satellite, NWP, solar geometry, and time calendar; and the U-Net decoder 235 is used to progressively recover spatial resolution by combining skip connections from the encoder.

[0083] Specifically, the noisy future scenario under the current diffusion step. Historical observation features and diffusion step encoding are first input into a 3D convolutional input module to obtain low-level spatiotemporal features. These low-level spatiotemporal features are further input into a U-Net encoder to obtain multi-scale encoded features and bottleneck features. Bottleneck features are input into a temporal Transformer module to model cross-temporal dependencies within the prediction window. Exogenous conditional features such as satellite, NWP, solar geometry, and time calendar are processed by a conditional encoder to obtain conditional features, which are then fused with bottleneck features or multi-scale encoded features through a cross-attention module. The fused features are input into a U-Net decoder and, combined with skip connections from the encoder, progressively restore spatial resolution. Finally, the conditional diffusion output head 240 outputs noise predictions, clean sample predictions, or velocity variable predictions.

[0084] The basic loss function is:

[0085] ,

[0086] Among them, L base Based on the total training loss; L diff For diffusion denoising loss; L energy The Energy Score loss is used to evaluate the overall distributional difference between the multi-scenario prediction ensemble and the actual future observations; L variogram The VariogramScore loss is used to evaluate the prediction quality of dependencies between different locations in a multivariate or spatial scene; L l1 The point-state L1 loss is used to constrain local numerical errors; w es w vs w l1 These are the weight parameters for the corresponding loss term.

[0087] When using noise prediction, the diffusion denoising loss can be expressed as:

[0088] ,

[0089] in, For diffusion denoising loss; This represents the real future scene without noise, that is, the real future solar irradiance field or photovoltaic power sequence in the training samples; This represents the model input conditions, including historical observation data and exogenous features; This indicates the diffusion step or noise level, which is different from the physical time step within the prediction window; This represents noise sampled from a standard Gaussian distribution; Indicates the first A diffusion step from a real future scenario Samples obtained by adding noise; The parameter is Conditional denoising networks on input noisy samples diffusion step and conditional input The noise obtained after prediction; The squared norm (L2) is used to measure true noise. With prediction noise The differences between them; Represents the real future scenarios in the training samples diffusion step and random noise Calculate the expected value. This objective enables the basic diffusion model to possess stable conditional sampling capabilities, future scene distribution modeling capabilities, and local numerical prediction capabilities.

[0090] Step S3: Freeze the reference model

[0091] Once the base model reaches the preset convergence condition, the base model will be... Copy and freeze as reference model f ref :

[0092] ,

[0093] Among them, f ref The parameters of the reference model remain unchanged during the post-training phase. The purpose of setting a reference model is to prevent the model from deviating completely from the base probability distribution, even though the sampling bias of the current model will be adjusted based on scheduling efficiency during post-training. By constraining the reference model, deviations between the current model and the base model in scene mean, scene distribution, latent representation, or denoising output can be limited, thereby reducing distribution drift and reward hacking risks.

[0094] Step S4: Generate candidate scenes based on conditional diffusion sampling

[0095] For each input X, the current model... Generate or select K candidate future scenarios:

[0096] ,

[0097] Where K is the number of candidate scenes, which can be 4, 8, 16 or other values.

[0098] In one implementation, candidate scenes are conditionally sampled from different initial noises using a DDPM, DDIM, DPM-Solver, or other backdiffusion sampler. That is, for the same input X, multiple possible future solar irradiance fields or photovoltaic power sequences are generated from different random noises through a backdiffusion process. K can take values ​​of 4, 8, 16, or other values. In a preferred implementation, K=8 to balance sampling diversity and training cost.

[0099] Step S5: Calculate the utility of multi-component decision-making

[0100] For each candidate scenario Y (k) Calculate the decision utility score:

[0101] ,

[0102] Where r k Y is the decision utility score for the k-th candidate scenario; U(·) is the decision utility function; gt This is for observing the real future. The decision utility function in this invention can essentially be understood as the reward function in reinforcement learning methods, that is, a comprehensive scoring function used to evaluate whether candidate scenarios are valuable for downstream scheduling decisions.

[0103] In one implementation, the decision utility function can be composed of a weighted average of multiple utility terms:

[0104] ,

[0105] in, The mass term of the basic distribution; For the hill climbing event item; This is an asymmetric underpredictable risk item; For reliability items during critical periods; This is the scheduling cost item; The weights of each utility item. In a preferred embodiment:

[0106] ,

[0107] The above weights can be adjusted based on the target power system, energy storage configuration, market rules, risk appetite, or operating scenario.

[0108] To avoid inconsistencies in the dimensions of different utility items, each utility item can be normalized before calculating the total utility. For example:

[0109] ,

[0110] Where m represents different utility terms, and These represent the mean and standard deviation of the utility term in the current batch or the historical sliding window, respectively, with ε being a small constant to prevent division by zero.

[0111] 1. Basic distribution mass term

[0112] The basic distribution quality term is used to constrain the overall similarity between candidate scenarios and real future observations, preventing the model from pursuing only scheduling rewards while ignoring basic prediction quality.

[0113] Assume there are a total of At each time step, the spatial grid set of the target region is... The number of grid points is Candidate scenarios at time step Spatial location The predicted value is denoted as The actual observed value is recorded as .

[0114] In one implementation, the mean absolute error between the candidate scenario and the actual future observation is first calculated:

[0115] ,

[0116] Then its opposite is used as the mass term of the basic distribution:

[0117] ,

[0118] Therefore, the closer the candidate scenario is to the actual observation, The smaller, The larger.

[0119] Alternatively, the L2 distance, weighted spatial distance, the negative of the Energy Score, or the Variogram Score can be used as the basic distribution quality term.

[0120] 2. Climbing event item

[0121] The ramp-up event item is used to evaluate whether candidate scenarios can identify rapid increases and decreases in solar irradiance or photovoltaic power.

[0122] First, the candidate region scenes are aggregated into a region average sequence:

[0123] ,

[0124] Correspondingly, real future observations are also aggregated into a regional average series:

[0125] ,

[0126] Then, the changes in adjacent time steps for the candidate scene and the actual observation are calculated separately:

[0127] ,

[0128] in, Indicates from arrive The change in regional average irradiance or photovoltaic power.

[0129] To ensure continuous calculation of hill-climb event indicators, a sigmoid function is used to construct soft hill-climb event indicators. For uphill events, the soft event value for candidate scenarios is defined as:

[0130] ,

[0131] The actual observed soft tags for uphill climbing are:

[0132] ,

[0133] in: For candidate scenarios at time steps Soft prediction of an uphill event; For accurate observation at time step Soft label indicating an uphill incident; This is the uphill threshold; The smoothing coefficient controls the transition speed of the sigmoid function from 0 to 1; This refers to the sigmoid function.

[0134] For downhill climbing events, the soft event value for candidate scenarios is defined as:

[0135] ,

[0136] The actual observed downhill soft tags are:

[0137] ,

[0138] in, This is the downhill threshold. Because downhill corresponds to... Significantly negative, therefore used here Indicates the magnitude of the decrease.

[0139] Next, for any type of climb... According to candidate soft prediction and real soft labels Calculate soft-TP, soft-FP, and soft-FN:

[0140] ,

[0141] in, To correctly identify climbing events in candidate scenarios, soft-true positives (soft-TPs) are used. Soft-FP is used to identify false positives for hill-climbing events in candidate scenarios. This is a soft-false negative (soft-FN) for candidate scenarios that missed reporting of hill-climbing events. The soft-F1 score for the hill-climbing event is then obtained.

[0142] ,

[0143] in, To prevent small constants with zero denominators, the soft-F1 scores for both uphill and downhill climbs are averaged and used as the hill-climb event recognition utility term for candidate scenes.

[0144] ,

[0145] Therefore, the more accurately candidate scenarios can identify rapid rise and rapid fall events in the real future, The larger.

[0146] 3. Asymmetric risk items

[0147] The asymmetric under-prediction risk term is used to reflect the different risks of under-prediction and over-prediction in the power system. For solar irradiance or photovoltaic power prediction tasks, under-prediction and over-prediction may correspond to different reserve capacity configurations, energy storage dispatch, or market deviation costs, so different penalty weights can be set.

[0148] First, define the prediction error for the candidate scene:

[0149] ,

[0150] when When the candidate scenario is lower than the actual observation, it indicates that under-prediction has occurred; when When the time is higher than the actual observation, it means that the prediction has occurred.

[0151] Define under-prediction magnitude and over-prediction magnitude:

[0152] ,

[0153] Then calculate the asymmetric prediction risk loss:

[0154] ,

[0155] in: Weighting for under-prediction penalties; For overprediction penalty weights; in one implementation, a setting can be provided. ,For example The opposite of this risk loss is taken as the utility term:

[0156] ,

[0157] Therefore, the lower the underprediction risk of the candidate scenario, especially when the underprediction penalty weight is high, its... The larger.

[0158] 4. Reliability items during critical periods

[0159] The critical period reliability item is used to emphasize the forecast quality during periods of high irradiance, high photovoltaic output, or high system value.

[0160] First, define the key time period mask. Taking high irradiance periods as an example, a threshold can be constructed based on the daylight periods in actual future observations:

[0161] ,

[0162] in: Thresholds for critical time periods; For daylight time periods; This represents the average value of the region as observed in the true future. It is the 70th percentile.

[0163] Define the key time period mask:

[0164] ,

[0165] in, Indicates time step This is a critical period. This indicates that it is not a critical period.

[0166] Then, the average prediction error of the candidate scene during the critical time period is calculated:

[0167] ,

[0168] Finally, its opposite is used as the reliability term for the critical period:

[0169] ,

[0170] Therefore, the smaller the error of the candidate scene during periods of high irradiance, high output, or high value, The larger.

[0171] 5. Scheduling cost item

[0172] The scheduling cost term is used to evaluate the actual value of candidate forecasting scenarios for energy storage dispatch, reserve capacity allocation, or electricity market bidding. Firstly, based on candidate forecasting scenarios... Generate a scheduling plan:

[0173] ,

[0174] in: According to the first A scheduling plan developed for each candidate prediction scenario; A set of feasible scheduling plans; In predicting scenarios The planned cost is evaluated below. In one embodiment, the scheduling plan can be calculated using linear programming, quadratic programming, dynamic programming, rule-based scheduling models, or neural network surrogate models. Linear programming, quadratic programming, or dynamic programming are used to solve the scheduling plan under constraints such as energy storage power, state of charge, reserve capacity, electricity price, or equipment operation; rule-based scheduling models are used to generate scheduling plans based on preset operating rules; and neural network surrogate models are used to approximate the actual scheduling optimization process or operating cost evaluation process.

[0175] Then, in real future observation The following is an assessment of the actual operating cost of this scheduling plan:

[0176] ,

[0177] Simultaneously, calculate the ex-post optimal baseline cost obtained directly from the true future results under the same scheduling constraints:

[0178] ,

[0179] Therefore, the first is defined Scheduling regret values ​​for each candidate scenario:

[0180] ,

[0181] Finally, the negative of the scheduling regret value is used as the scheduling utility term:

[0182] ,

[0183] Therefore, the closer the scheduling plan corresponding to the candidate scenario is to Oracle's optimal scheduling, the smaller its regret value. The larger.

[0184] If an explicit scheduling optimization model is not available for a particular application scenario, alternative methods such as standby capacity cost, energy storage charging and discharging proxy model, market deviation penalty model, or neural network proxy can be used. and The alternative implementation.

[0185] Step S6: Construct the relative target distribution within the group

[0186] The utility scores {r1, r2, …, r} of candidate scenes under the same input X are used to calculate the utility scores of the scenes. k The distribution is transformed into the target scene distribution using a softmax function with a temperature parameter.

[0187] ,

[0188] Where, q k Let be the probability of the k-th candidate scene in the target distribution; is the utility temperature parameter, used to control the sharpness of the target distribution; K is the number of candidate scenes. When When the target is smaller, the target distribution is more biased towards the candidate scenario with the highest utility; when When the target size is larger, the target distribution is smoother. In one implementation, = 0.2.

[0189] Multiple candidate scenes generated under the same input conditions are treated as a single group, and a list-based target distribution is constructed based on the relative decision utility of the candidate scenes. This target distribution indicates that, under the same input conditions, candidate scenes with higher utility should receive a higher generation probability.

[0190] Step S7: Perform listwise decision alignment diffusion followed by training

[0191] See Figure 3 For candidate scenario Y (k) Calculate the approximate generated score under the current diffusion model:

[0192] ,

[0193] in, For the current model, the candidate scene Y (k) The generated score; For the current model, the candidate scene Y (k) The diffusion denoising loss is smaller if the current diffusion model can better interpret or reconstruct a candidate scene, resulting in a higher generation score.

[0194] Alternatively, diffusion ELBO, denoised trajectory likelihood approximation, score matching energy, or sampling path score can be used as... Then, the candidate scene scores are converted into a model-induced distribution:

[0195] ,

[0196] in, The relative probability of the k-th candidate scene induced by the current model; This is the temperature parameter for the model distribution, used to control the smoothness of the model-induced distribution. Optimize the current model to improve the smoothness of the model-induced distribution. The target distribution q is approximated by the decision utility. The within-group target distribution q can be constructed using softmax, rank-based normalization, top-k normalization, or advantage normalization. In one implementation, the listwise decision alignment loss is:

[0197] ,

[0198] in, Losses due to decision alignment; For the target distribution q and the model-induced distribution The KL divergence between them; α is the weight parameter of the weighted diffusion denoising loss; the second term is used to maintain the stability of the diffusion model's denoising training and avoid destroying the generation quality by only optimizing the relative ranking.

[0199] The core function of this step is that, unlike single-sample error optimization or pairwise winner / loser comparison, this invention treats multiple candidate scenarios under the same input conditions as a whole, constructs a listwise target distribution based on their relative decision utility, and makes the model's generation tendency to move closer to this target distribution, thereby directly changing the sampling preference of the diffusion model.

[0200] Step S8: Add stability constraints

[0201] To avoid distribution drift, physically unreasonable outputs, or probability calibration degradation caused by the model excessively pursuing decision utility during post-training, this invention adds stability constraints in addition to decision alignment loss.

[0202] The total post-training loss is:

[0203] ,

[0204] Where: L is the total post-training loss; L align For list-based decision alignment loss; L ref For reference model consistency constraints; L phys For physical consistency constraints; L cal For probability calibration constraints; These are the weight parameters for each constraint term.

[0205] 1. Reference Model Consistency Constraints

[0206] Reference model consistency constraints are used to limit the current model relative to the frozen reference model. The deviation. This constraint can be applied to the scene mean, scene probability, denoised output, or implicit representation.

[0207] In one implementation, noise reduction output consistency can be employed:

[0208] ,

[0209] Among them, among them, Consistency constraints for the reference model; Indicates the first Noisy samples under each diffusion step; Indicates the diffusion step or noise level; This represents the model input conditions, including historical observation data and exogenous features; This indicates that the current model is using noisy input samples. diffusion step and conditional input The resulting denoised output; This represents the denoised output obtained by freezing the reference model under the same input; Indicates the addition of noise to the sample diffusion step and input conditions Seeking expectations; represents the squared L2 norm, used to measure the difference between the denoised output of the current model and the denoised output of the frozen reference model. By minimizing this constraint, the current model can be limited from excessively shifting relative to the base model in the post-training stage, thereby reducing the risk of distribution drift and degradation in generation quality due to excessive pursuit of decision utility.

[0210] 2. Physical consistency constraints

[0211] Physical consistency constraints are used to ensure that the predicted scenario meets the basic physical laws of solar irradiance or photovoltaic power, including nonnegativity, near-zero at night, temporal smoothness, and spatial continuity.

[0212] In one implementation:

[0213] ,

[0214] Among them, L nonneg This is a non-negativity constraint used to penalize irradiance or power less than zero; L night The near-zero constraint at night is used to constrain predicted values ​​to be close to zero during nighttime or low solar altitude angle periods; L smoothFor time smoothness constraints, used to avoid unreasonable time jumps; L tv This is a regularization of the total spatial variation, used to maintain the continuity of the spatial scene.

[0215] For example, a nonnegativity constraint can be expressed as:

[0216] ,

[0217] Physical consistency constraints can be extended to cloud motion continuity constraints, clear sky upper bound constraints, spatial gradient constraints, or capacity constraints, depending on the application.

[0218] 3. Probability calibration constraints

[0219] Probabilistic calibration constraints are used to ensure that the uncertainty of the model output matches the true error, avoiding excessively wide or narrow prediction intervals or distorted coverage after training.

[0220] In one implementation, interval coverage error can be used:

[0221] in: To predict interval coverage; The target coverage rate is, for example, 90% or 95%. Alternatively, the PIT statistic, reliability penalty, or quantile calibration error can be used as probabilistic calibration constraints.

[0222] In one implementation, the total post-training loss can be specifically written as:

[0223] ,

[0224] The weights mentioned above are for illustrative purposes only and can be adjusted based on the specific dataset, target region, prediction task, and downstream scheduling requirements.

[0225] Step S9: Output the probability prediction results after decision alignment

[0226] After training, the model outputs one or more results from a set of scenarios for any input X: multiple future solar irradiance scenarios, probability weights for each scenario, weighted mean prediction, prediction intervals or quantiles, and scenarios available for energy storage scheduling, reserve capacity configuration, ramp-up event warnings, or electricity market bidding.

[0227] Example 1

[0228] This embodiment uses the solar irradiance diffusion scenario prediction model trained by the above post-training method for regional-level solar irradiance scenario prediction. The specific steps are as follows:

[0229] Acquire historical GHI and GOES satellite cloud images, solar altitude angle, solar azimuth angle, hourly data, and daily data for the target area from 2018 to 2023.

[0230] Convert GHI to clear-sky index and clip the region to a 64 x 64 grid.

[0231] The past 4 hours are used as the input window, and the next 4 hours are used as the output window, with a time interval of 15 minutes.

[0232] A conditional diffusion model is constructed, and its denoising network adopts the Hybrid ST-UNet structure, with Temporal Transformer and cross-attention fusion module introduced in the bottleneck layer.

[0233] The base model is trained using diffusion denoising loss and combined with Energy Score + VariogramScore + pointwise L1 based on a sampled scene set.

[0234] The frozen base model is used as a reference model.

[0235] In the post-training phase, for each input, eight candidate scenes are sampled starting from different initial noise, and five rewards are calculated to obtain the target distribution q within the group.

[0236] pi_theta is constructed using the approximate generated scores of candidate scenes under the current diffusion model, and post-training is performed based on KL(q || pi_theta), weighted diffusion denoising loss, reference model constraints, physical constraints, and calibration constraints.

[0237] The output of multi-scenario prediction results after training is used for hill climbing event identification and scheduling value evaluation.

[0238] In this embodiment, the base model is a conditional diffusion model. Its probabilistic nature comes from the backsampling process of the future irradiance field conditional distribution, which can naturally generate multiple future scenes with spatial continuity and temporal consistency. If the deployment environment has high requirements for inference latency, DDIM, DPM-Solver, distillation sampling, or few-step sampling methods can be used to reduce sampling costs.

[0239] Example 2

[0240] This embodiment uses the solar irradiance diffusion scenario prediction model trained by the above post-training method for site-level PV power prediction. The specific steps are as follows:

[0241] The region model encoder trained in Example 1 is used as a feature extractor.

[0242] Extract the implicit representation of the corresponding spatial location or neighborhood of the target photovoltaic site.

[0243] Replace the output head with the site power scene output head to output the normalized PV power scene for multiple future time steps.

[0244] The post-training utility function retains the quality of the base distribution, ramp events, asymmetric underprediction risk, and critical time period items; it selects whether to enable R_dispatch based on whether a scheduling model is available.

[0245] Lightweight adjustments were made using historical power data from the site.

[0246] This embodiment demonstrates that the present invention is applicable not only to regional irradiance fields but also to site-level photovoltaic power prediction.

[0247] Example 3

[0248] This embodiment applies the solar irradiance diffusion scenario prediction model trained by the above post-training method to photovoltaic power plants or aggregators with energy storage systems. The specific steps are as follows:

[0249] For each candidate prediction scenario, input the energy storage scheduling agent model.

[0250] The energy storage scheduling agent model outputs the charging and discharging plans and expected operating costs corresponding to the predicted scenarios.

[0251] Oracle dispatch cost is calculated using the true future power curve.

[0252] The difference between the predicted dispatch cost and the oracle dispatch cost is defined as dispatch regret.

[0253] Add the negative of dispatch regret as R_dispatch to the total utility function.

[0254] Increase the probability of generating low-regret scenarios during the post-training phase.

[0255] This embodiment enables the prediction model to directly learn the scenario distributions that are more valuable for energy storage scheduling, rather than just learning prediction results with smaller average errors.

[0256] Those skilled in the art should understand that variations can be implemented by combining existing technology with the above embodiments, which will not be elaborated here. Such variations do not affect the essence of the present invention, and will not be elaborated here either.

[0257] The preferred embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above. Systems and structures not described in detail should be understood as being implemented in a conventional manner in the art. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the present invention. This does not affect the essential content of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the present invention are still within the scope of protection of the present invention.

Claims

1. A post-training method for a solar irradiance diffusion scenario prediction model, characterized in that: Includes the following steps: The conditional diffusion model, which has been trained on a basic basis, is used as a basic probabilistic scene generator to generate multiple future scenes; after the conditional diffusion model is trained on a basic basis, the basic model parameters are frozen and kept from being updated in the later training stage. Calculating multi-component decision utility: The decision utility function is composed of a weighted average of multiple utility terms. in, The mass term of the basic distribution; For the hill climbing event item; This is an asymmetric underpredictable risk item; For reliability items during critical periods; This is the scheduling cost item; The weights of each utility item; Multiple candidate scenarios generated under the same input conditions are treated as a whole within a group, and the relative target distribution within the group is constructed based on the relative decision utility of the candidate scenarios. Post-training after performing list-based decision alignment diffusion: In addition to the decision alignment loss, reference model constraints, physical consistency constraints, and probabilistic calibration constraints are introduced. The total loss function after post-training is: Where: L is the total post-training loss; L align For list-based decision alignment loss; L ref For reference model consistency constraints; L phys For physical consistency constraints; L cal For probability calibration constraints; These are the weight parameters for each constraint term; Determined through the following process: for any type of climb... According to candidate soft prediction and real soft labels Calculate soft-TP, soft-FP, and soft-FN: in, To correctly identify climbing events in candidate scenarios, soft-true positives (soft-TPs) are used. Soft-FP is used to identify false positives for hill-climbing events in candidate scenarios. The soft-FN is used to identify the false negatives for missed climbing events in candidate scenarios; thus, the soft-F1 score for the climbing events is obtained. in, To prevent small constants with zero denominators, the soft-F1 scores for both uphill and downhill climbs are averaged and used as the hill-climbing event recognition utility term for candidate scenes. .

2. The post-training method for a solar irradiance diffusion scenario prediction model according to claim 1, characterized in that, The conditional diffusion model includes a forward noise addition process and a reverse noise reduction process: First, the noisy future scene, historical observation features, and diffusion step encoding under the current diffusion step are input into a 3D convolutional input module to obtain low-level spatiotemporal features; the low-level spatiotemporal features are further input into a U-Net encoder to obtain multi-scale encoded features and bottleneck features; the bottleneck features are input into a temporal Transformer module to model cross-temporal dependencies within the prediction window; Exogenous conditional features, including satellite, NWP, solar geometry, and time calendar, are processed by a conditional encoder to obtain conditional features, which are then fused with bottleneck features or multi-scale coded features through a cross-attention module. The fused features are input into the U-Net decoder and, combined with skip connections from the encoder, are used to restore spatial resolution step by step. Finally, a conditional diffusion denoising head outputs noise prediction, clean sample prediction, or velocity variable prediction.

3. The post-training method for a solar irradiance diffusion scenario prediction model according to claim 1, characterized in that, The loss function for the basic training of the conditional diffusion model is: Among them, L base Based on the total training loss; L diff For diffusion denoising loss; L energy The Energy Score loss is used to evaluate the overall distributional difference between the multi-scenario prediction ensemble and the actual future observations; L variogram Variogram Score loss is used to evaluate the prediction quality of the dependency structure between different locations in a multivariate or spatial scene; L l1 The point-state L1 loss is used to constrain local numerical errors; w es w vs w l1 These are the weight parameters for the corresponding loss term.

4. The post-training method for a solar irradiance diffusion scenario prediction model according to claim 1, characterized in that, The basic distribution quality term constrains the overall similarity between candidate scenarios and actual future observations; the ramp-up event term evaluates whether candidate scenarios can identify rapid increases and decreases in solar irradiance or photovoltaic power; the asymmetric underprediction risk term evaluates the different risks of underprediction and overprediction in the power system; and the critical period reliability term evaluates the prediction quality during periods of high irradiance, high photovoltaic output, or high system value. The evaluation of the scheduling cost item assesses the actual value of candidate prediction scenarios for energy storage scheduling, reserve capacity configuration, or electricity market bidding.

5. The post-training method for a solar irradiance diffusion scenario prediction model according to claim 4, characterized in that, The basic distribution quality term is the negative of the mean absolute error between the candidate scenario and the actual future observation; the climbing event term is the average of the soft-F1 scores of the uphill and downhill climbs; the asymmetric underprediction risk term is the negative of the asymmetric prediction risk loss. Where: T is the number of time steps within the prediction window, For spatial grid point set or power station set, For time indexing, Weighting for under-prediction penalties; Over-prediction penalty weights; The magnitude is under-predicted; The over-prediction magnitude is defined as follows: the critical period reliability term is the negative of the average prediction error of the candidate scenario during the critical period; the scheduling cost term is the negative of the scheduling regret value, which is the difference between the actual operating cost of the scheduling plan generated based on the candidate prediction scenario under the real future result and the ex-post optimal cost obtained by directly solving based on the real future result under the same scheduling constraints.

6. The post-training method for a solar irradiance diffusion scenario prediction model according to claim 1, characterized in that, Constructing the relative target distribution within a group includes: dividing the candidate scene utility scores {r1, r2, …, r} under the same input X. k The distribution is transformed into the target scene distribution using a softmax function with a temperature parameter. Where, q k Let be the probability of the k-th candidate scene in the target distribution; is the utility temperature parameter, used to control the sharpness of the target distribution; K is the number of candidate scenes.

7. The post-training method for a solar irradiance diffusion scenario prediction model according to claim 1, characterized in that, The list-based decision alignment loss is: in, Losses due to decision alignment; For the target distribution q and the model-induced distribution The KL divergence between them; α is the weighting parameter of the weighted diffusion denoising loss; For the current model, the candidate scene Y (k) The diffusion denoising loss; The reference model consistency constraint limits the deviation of the current model from the frozen reference model; the physical consistency constraint ensures that the predicted scenario meets the basic physical laws of solar irradiance or photovoltaic power; and the probabilistic calibration constraint ensures that the uncertainty of the model output matches the actual error.

8. A solar irradiance diffusion scenario prediction model, characterized in that, The conditional diffusion model is obtained by training the base training completed using the post-training method as described in any one of claims 1 to 7.

9. The solar irradiance diffusion scenario prediction model according to claim 8, characterized in that, The solar irradiance diffusion scenario prediction model outputs one or more results from a set of scenarios for any input X, including multiple future solar irradiance scenarios, probability weights for each scenario, weighted mean prediction, prediction intervals or quantiles, and scenarios available for energy storage dispatch, reserve capacity configuration, ramp-up event warning, or electricity market bidding.

10. A computer program product, characterized in that, When the program runs on the processor, it executes the post-training method of the solar irradiance diffusion scenario prediction model as described in any one of claims 1 to 7.

Images