Wind-solar-hydro multi-energy hybrid system source-side key time series interval forecasting and credibility evaluation method

Abstract

The application discloses a kind of wind light water multi-energy hybrid system source side key time sequence interval prediction and credibility evaluation method, comprising: obtaining multidimensional time sequence input sample set;Causal structure model is constructed to decouple exogenous natural output and scheduling control effect, obtain exogenous uncertainty to be calibrated data and control effect to be evaluated data;Only for the exogenous uncertainty to be calibrated data, multi-time scale conforming calibration is carried out, and exogenous uncertainty calibration result is obtained;Finally, the exogenous uncertainty calibration result, the control effect to be evaluated data, the pre-stored physical constraint and operation boundary model, and the scheduling decision demand are comprehensively generated, and the multi-energy interval prediction result that is physically feasible and decision-related is generated.The application solves the problem of error source confusion and statistical and physical disconnection in traditional prediction, and improves the robustness and availability of interval prediction under the change of scheduling strategy.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of multi-energy complementary technology, and particularly relates to a method for source-side key time sequence interval prediction and credibility evaluation of a wind-solar-hydro multi-energy hybrid system. BACKGROUND

[0002] With high-proportion renewable energy such as wind power, photovoltaic power and hydroelectric power being connected, a wind-solar-hydro multi-energy hybrid system has become an important part of power system dispatching. How to accurately predict the output of these multi-energies with strong uncertainty and evaluate the credibility thereof has great research significance for guaranteeing safe operation of a power grid, improving dispatching efficiency and promoting clean energy consumption.

[0003] At present, researches on output prediction of a wind-solar-hydro hybrid system mainly focus on using a deep learning model (such as a recurrent neural network or a long short-term memory network) to improve the accuracy of point prediction. In the aspect of interval prediction, methods include quantile regression, kernel density estimation and ensemble learning. In recent years, conformal prediction technology has also been used to generate a prediction interval with strict statistical coverage rate guarantee for a point prediction result, so as to quantify the uncertainty of prediction.

[0004] The existing technology mainly faces the core problems of confusion of prediction error sources and disconnection of statistical intervals and physical constraints. First, the traditional method models and calibrates the total output error as a whole and fails to distinguish between uncontrollable exogenous natural fluctuations (such as meteorological mutations) and controllable dispatching control effects (such as artificial power cuts). This leads to the fact that when the dispatching strategy changes, the prediction interval constructed based on the historical total error distribution will quickly fail. Secondly, the generated prediction interval only meets the statistical coverage rate, but may contain a large number of trajectories that violate the physical constraints of the power grid (such as power flow or climbing limit), and cannot be directly adopted by the dispatching system, resulting in a huge gap between statistical credibility and physical feasibility. SUMMARY

[0005] The application aims to provide a method for source-side key time sequence interval prediction and credibility evaluation of a wind-solar-hydro multi-energy hybrid system, so as to solve one of the problems existing in the prior art.

[0006] The technical scheme is a method for source-side key time sequence interval prediction and credibility evaluation of a wind-solar-hydro multi-energy hybrid system, comprising the following steps:

[0007] acquiring a multi-dimensional time sequence input sample set;

[0008] based on the multi-dimensional time sequence input sample set, constructing a causal structure model to decouple exogenous natural output and dispatching control effect, to obtain exogenous uncertainty to be calibrated data and control effect to be evaluated data;

[0009] only for the exogenous uncertainty to be calibrated data, performing multi-time scale conformal calibration to obtain an exogenous uncertainty calibration result;

[0010] The exogenous uncertainty calibration result, the control effect to be evaluated data, the pre-stored physical constraint and operation boundary model, and the scheduling decision demand are comprehensively integrated to generate a multi-energy interval prediction result.

[0011] According to an aspect of the present application, the step of constructing a causal structure model to decouple exogenous natural output and scheduling control effect to obtain exogenous uncertainty to be calibrated data and control effect to be evaluated data includes:

[0012] constructing a causal structure model;

[0013] constructing an exogenous natural output prediction model according to the causal structure model, and generating an exogenous natural output time sequence and an exogenous natural output prediction time sequence;

[0014] constructing a control effect time sequence, and constructing a control effect prediction model to generate a control effect prediction time sequence;

[0015] calculating an exogenous residual time sequence based on the difference between the exogenous natural output time sequence and the exogenous natural output prediction time sequence, and combining the exogenous natural output prediction time sequence to form the exogenous uncertainty to be calibrated data;

[0016] calculating a control residual time sequence based on the difference between the control effect time sequence and the control effect prediction time sequence, and combining the control effect prediction time sequence to form the control effect to be evaluated data.

[0017] According to an aspect of the present application, the step of constructing an exogenous natural output prediction model and generating an exogenous natural output time sequence and an exogenous natural output prediction time sequence includes:

[0018] determining a benchmark control determination rule, the rule being used to identify intervals close to no control intervention in the multi-dimensional time sequence input sample set;

[0019] applying the benchmark control determination rule to screen the multi-dimensional time sequence input sample set to obtain a benchmark control sample set;

[0020] specifying the historical multi-energy output time sequence contained in the benchmark control sample set as the exogenous natural output time sequence;

[0021] training the exogenous natural output prediction model only by using the benchmark control sample set.

[0022] According to an aspect of the present application, the step of constructing a control effect time sequence and constructing a control effect prediction model to generate a control effect prediction time sequence includes:

[0023] Historical multi-energy output time series are extracted from the multi-dimensional time series input sample set, and the historical multi-energy output time series are differiated with the exogenous natural output prediction time series to construct the control effect time series;

[0024] Extract the time series of unit and reservoir status and the time series of scheduling and control actions from the multi-dimensional time series input sample set to construct the control effect input feature sequence;

[0025] A control effect prediction model is trained by using the control effect input feature sequence as the model input and the control effect time series as the model target.

[0026] According to one aspect of this application, the steps for constructing a causal structure model include:

[0027] Based on the physical mechanism and operational experience of the wind-solar-hydro multi-energy hybrid system, an initial causal structure model is set up, which explicitly distinguishes between exogenous driving paths and scheduling control paths.

[0028] Based on a multi-dimensional time-series input sample set, a data-driven causal discovery method is used to revise and verify the initial causal structure model, resulting in a causal structure model.

[0029] According to one aspect of this application, the steps for performing multi-timescale conformal calibration to obtain exogenous uncertainty calibration results include:

[0030] Determine the set of time scales that are appropriate for scheduling requirements;

[0031] By utilizing the exogenous residual time series in the data to be calibrated with exogenous uncertainty, calibration samples are constructed at each time scale in the time scale set, and nonconformal scores are calculated to obtain a multi-time scale nonconformal score set.

[0032] Based on the multi-timescale nonconformal score set, the quantile corrections at each time scale are obtained at a given confidence level, generating a multi-timescale quantile correction set.

[0033] By fusing a set of quantile corrections across multiple time scales, an exogenous interval correction sequence is generated.

[0034] By combining the exogenous interval correction sequence with the exogenous natural power output prediction time series in the exogenous uncertainty data to be calibrated, the exogenous natural power output interval forecast results are constructed.

[0035] Among them, the exogenous uncertainty calibration results include the exogenous interval correction sequence and the exogenous natural output interval prediction results.

[0036] According to one aspect of this application, the step of fusing a set of multi-timescale quantile corrections to generate an exogenous interval correction sequence includes:

[0037] setting a time scale fusion weight;

[0038] on the validation set, a performance index function is constructed to target coverage bias and interval width;

[0039] by optimizing the performance index function, the time scale fusion weight is determined;

[0040] the time scale fusion weight is applied to the weighted combination of the multi-time scale quantile correction quantity set to generate the exogenous interval correction quantity sequence.

[0041] According to one aspect of the present application, the steps of generating a multi-energy interval prediction result based on the exogenous uncertainty calibration result, the control effect to be evaluated data, the physical constraint and operation boundary model, and the scheduling decision demand include:

[0042] Combine the exogenous natural output interval prediction result in the exogenous uncertainty calibration result with the control effect prediction time series in the control effect to be evaluated data to generate a candidate multi-energy output trajectory set;

[0043] Inject the candidate multi-energy output trajectory set into the physical constraint and operation boundary model to evaluate the physical constraint violation degree of the candidate multi-energy output trajectory set, and obtain a physical constraint violation degree index set;

[0044] Remove trajectories in the candidate multi-energy output trajectory set whose physical constraint violation degree exceeds a preset threshold to form a physically feasible trajectory set;

[0045] And based on the physically feasible trajectory set, the physical constraint violation degree index set, the exogenous interval correction quantity sequence in the exogenous uncertainty calibration result, the control residual time series in the control effect to be evaluated data, and the scheduling decision demand, a multi-energy interval prediction result is generated.

[0046] According to one aspect of the present application, the steps of generating a candidate multi-energy output trajectory set include:

[0047] A candidate control strategy set is constructed;

[0048] A representative exogenous trajectory set is extracted from the exogenous natural output interval prediction result;

[0049] For any representative exogenous trajectory in the representative exogenous trajectory set, and in combination with any candidate control strategy in the candidate control strategy set, the control effect prediction time series is adjusted to obtain a control effect adjustment time series;

[0050] The control effect adjustment time series and the representative exogenous trajectory are added point by point to construct a candidate multi-energy output trajectory, and further form a candidate multi-energy output trajectory set.

[0051] According to one aspect of the present application, based on the set of physically feasible trajectories, the set of physical constraint violation indicators, the sequence of exogenous interval correction amounts, the time series of control residuals, and the set of scheduling decision requirement descriptions, the step of generating the multi-energy interval prediction result further comprises:

[0052] Based on the set of physically feasible trajectories, the set of physical constraint violation indicators, the sequence of exogenous interval correction amounts, and the time series of control residuals, trajectory features are extracted.

[0053] The trajectory features are clustered to form a set of operating scenario clusters.

[0054] A comprehensive confidence score function is constructed, which is used to integrate the physical risk from the set of physical constraint violation indicators, the exogenous uncertainty from the sequence of exogenous interval correction amounts, and the control modeling error from the time series of control residuals.

[0055] The comprehensive confidence score function is applied to the set of physically feasible trajectories to generate a set of source-side comprehensive confidence evaluation indicators.

[0056] And according to the set of source-side comprehensive confidence evaluation indicators and the set of scheduling decision requirement descriptions, the multi-energy interval prediction result is generated.

[0057] The beneficial effects, the present application aims at the error source confusion problem, by constructing a causal structure model and a double-branch residual modeling, the total error is decomposed into uncontrollable exogenous residuals and controllable control residuals. By only performing conformal calibration on the exogenous residuals, the robustness of the prediction interval when the scheduling strategy changes is ensured. For the problem of disconnection between statistics and physics, by embedding the grid physical constraint model into the trajectory screening process, physically infeasible trajectories are removed; and by constructing a decision-aware comprehensive confidence score function, the physical risk, the exogenous uncertainty and the scheduling sensitivity are coupled, so that the final interval prediction result can accurately reflect the scheduling risk, solving the problem of disconnection between prediction and decision requirements. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 The flowchart of the present application as a whole.

[0059] Figure 2 The flowchart of the present application for obtaining exogenous uncertainty to be calibrated data and control effect to be evaluated data.

[0060] Figure 3 The flowchart of the present application for generating exogenous natural power time series and exogenous natural power prediction time series.

[0061] Figure 4 The flowchart of the present application for generating control effect prediction time series. DETAILED DESCRIPTION

[0062] Embodiment 1, a method for source-side key time series interval forecasting and credibility assessment of wind-solar-hydro multi-energy hybrid system is described. The method introduces a causal decoupling step before conformal calibration, distinguishing between uncontrollable exogenous uncertainty and controllable scheduling effect, thereby providing a more accurate basis for subsequent conformal calibration and credibility assessment.

[0063] Step 101, obtain a multi-dimensional time series input sample set.

[0064] In this embodiment, this step aims to build a unified basis sample set for subsequent causal modeling and conformal calibration. Specifically, this step first involves reading and structuring the multi-source raw data. For example, from the monitoring system or database of the wind-solar-hydro multi-energy hybrid system, read the historical multi-energy output time series, meteorological element time series X_met, unit and reservoir state time series X_state, and scheduling and control action time series U_ctrl.

[0065] The historical multi-energy output time series includes wind power output historical time series Y_hist_wind, photovoltaic output historical time series Y_hist_pv, and hydroelectric output historical time series Y_hist_hydro, etc.

[0066] The scheduling and control action time series U_ctrl includes output instruction, power cut instruction, start-stop instruction, etc.

[0067] The meteorological element time series X_met includes wind speed, wind direction, irradiance, air temperature, etc.

[0068] On this basis, multi-source time alignment and uniform sampling step are performed. Since the original data comes from different sources, their sampling frequencies may not be consistent. Preferably, a unified target sampling interval is determined according to the scheduling requirements, for example, a fifteen-minute time step, and all the above original data are time-aligned and resampled. For missing time points that occur during the resampling process, time interpolation or nearest neighbor filling can be performed.

[0069] Anomaly detection and data cleaning are performed on the time-aligned sample set. This step is used to eliminate obviously erroneous data points. For example, a combination strategy based on physical constraint rules and statistical thresholds (such as 3-sigma rule) can be used to identify and correct abnormal values. Identified abnormal points can be reasonably interpolated or directly removed. Based on physical constraint rules include output value exceeding unit rated output range.

[0070] The numerical normalization and multi-dimensional sample set construction are performed. In order to eliminate the influence of different physical dimensions, the interval normalization or standardization processing is performed on the cleaned time series of each type, such as mapping to the 0-1 interval or Z-Score standardization. For classification features, such as scheduling instruction types, one-hot encoding or embedding encoding can be performed. Finally, all the processed features are spliced according to a unified time index to form a structured multi-dimensional time series input sample set D_clean.

[0071] In step 102, based on the multi-dimensional time series input sample set, a causal structure model is constructed to decouple the exogenous natural output and the scheduling control effect, to obtain exogenous uncertainty to be calibrated data and control effect to be evaluated data.

[0072] Since the existing conformal prediction method generally models the total output error as a whole, it fails to distinguish the source of the error. The total error contains both the exogenous fluctuations caused by uncontrollable factors such as weather, and the artificial deviation caused by scheduling control behaviors such as power limiting and start-stop. This step aims to explicitly separate the two effects by constructing a causal structure model.

[0073] In this embodiment, the exogenous uncertainty to be calibrated data preferably includes an exogenous residual time series E_exo and an exogenous natural output prediction time series Y_exo_hat. The control effect to be evaluated data preferably includes a control residual time series E_ctrl and a control effect prediction time series Y_ctrl_hat. The detailed implementation of this step, including how to construct a causal structure model, how to train an exogenous output model using a benchmark control sample set, and how to calculate exogenous residuals and control residuals respectively, will be described in detail in Embodiment 2.

[0074] In step 103, only for the exogenous uncertainty to be calibrated data, multi-time scale conformal calibration is performed to obtain the exogenous uncertainty calibration result.

[0075] This step only calibrates the exogenous uncertainty data. Specifically, the calibration process only uses the exogenous residual time series E_exo obtained in Embodiment 2 to calculate the non-conformal score, and the control residual time series E_ctrl does not participate in this step. The technical effect of this is that the statistical confidence guarantee provided by conformal calibration strictly corresponds to uncontrollable exogenous factors such as weather, and is not contaminated by controllable scheduling strategy changes.

[0076] In this embodiment, the multiple time scales refer to the calibration process considering multiple time scales adapted to scheduling requirements, such as fifteen minutes, one hour, and four hours. The exogenous uncertainty calibration result preferably includes the exogenous interval correction sequence Δ_exo and the finally constructed exogenous natural output interval prediction result Y_exo_interval. The detailed implementation of this step, including how to calculate the multi-scale nonconformal score, how to obtain the quantile correction, and how to fuse weights, will be described in detail in Embodiment 3.

[0077] Step 104: By combining the exogenous uncertainty calibration results, the control effect data to be evaluated, the pre-stored physical constraints and operational boundary models, and the scheduling decision requirements, multi-energy range forecast results are generated.

[0078] This step is the final stage of the method of the present invention, which aims to combine the statistical prediction results obtained from the preceding steps with the actual physical power grid constraints and scheduling objectives to generate final usable decision support information.

[0079] Specifically, the exogenous uncertainty calibration result, such as Y_exo_interval, provides the fluctuation range of exogenous natural output. The control effect evaluation data, such as Y_ctrl_hat, is used to superimpose different control strategies. The pre-stored physical constraint and operational boundary model M_grid is key to achieving statistical-physical dual feasibility. This model may specifically include grid power flow constraints, line carrying capacity, unit ramping constraints, and reserve constraints. The scheduling decision requirement D_sched_req represents the optimization objective of the scheduling center, such as economic objectives (e.g., generation cost, wind and solar curtailment rate) or safety objectives (e.g., load loss risk).

[0080] By integrating all the above information, this method first generates a set of candidate multi-energy output trajectories, then filters them using a physical constraint model, and next performs a decision-aware credibility assessment on the physically feasible trajectories. Finally, it filters or weights these trajectories to obtain multi-energy interval prediction results for scheduling. The detailed implementation of this step, including how to perform physical constraint embedded trajectory filtering, how to construct a comprehensive credibility scoring function, and how to generate the final interval, will be elaborated in Examples 4 and 5.

[0081] Example 2 describes a two-branch residual modeling based on causal decoupling, i.e., the causal decoupling process described in step 102 of Example 1. In other words, a two-branch prediction model is constructed, which models the exogenous natural output and scheduling control effects separately under the constraints of the causal structure model, thereby obtaining the exogenous residual and control residual required for subsequent steps.

[0082] Step 201, a causal structure model can be constructed based on the multi-dimensional time series input sample set.

[0083] The influence paths of different factors on the total output are distinguished in structure, as follows:

[0084] The roles of variables in the multi-dimensional time series input sample set D_clean need to be divided. For example, the meteorological element time series X_met is designated as an exogenous driving variable; the unit and reservoir state time series X_state and the scheduling and control action time series U_ctrl are designated as controllable or human decision variables; the historical multi-energy output time series Y_hist_multi is designated as an explained output variable.

[0085] As a preferred embodiment, this step can be implemented in combination with expert knowledge and data-driven methods. Specifically, first, in combination with the physical mechanism and operation experience of the wind-solar-hydro multi-energy hybrid system, an initial causal structure model G_causal_init is set. The initial model explicitly distinguishes between exogenous driving paths (e.g., from X_met to Y_hist_multi) and scheduling control paths (e.g., from U_ctrl and X_state to Y_hist_multi).

[0086] Based on the multi-dimensional time series input sample set, a data-driven causal discovery method is used to modify and verify the initial causal structure model. For example, it is verified whether some edges in the initial model exist significantly in the data or the missing significant correlations are supplemented, and finally a modified causal structure model G_causal that passes the consistency check (such as no causal cycle) is obtained. The causal discovery method includes methods based on conditional independence test or scoring search, such as PC algorithm, FCI algorithm, etc.

[0087] The causal structure model G_causal can be represented conceptually by a structural equation, for example: Y_hist_multi(t) = F_exo(X_met(t)) + F_ctrl(U_ctrl(t), X_state(t)) + ε(t), where F_exo represents the exogenous natural output mapping, and F_ctrl represents the control effect mapping.

[0088] Step 202, according to the causal structure model, an exogenous natural output prediction model is constructed, and an exogenous natural output time series and an exogenous natural output prediction time series are generated. The purpose is to model the exogenous natural output branch separately.

[0089] As a preferred embodiment, the specific implementation of this step includes:

[0090] A base control decision rule Rule_base is determined. The rule is used to identify those time windows from the multi-dimensional time series input sample set D_clean that are close to no-control intervention. For example, the rule can be defined as the time windows where the power curtailment instruction magnitude is below a pre-set threshold and the start-stop action is infrequent, the threshold can be set as 5% below the rated power. In other words, the base control is regular / light control rather than pure ideal case of zero control.

[0091] The base control sample set D_base is obtained by applying the base control decision rule to filter the multi-dimensional time series input sample set D_clean.

[0092] The historical multi-energy output time series Y_hist_multi contained in the base control sample set D_base is designated as the exogenous natural output time series Y_exo_true. This definition is reasonable because in the base period of no-control intervention, the actual output is equivalent to the exogenous natural output.

[0093] The exogenous natural output prediction model M_exo is trained only using the base control sample set D_base. Specifically, M_exo is trained with the meteorological element time series X_met in D_base as model input and the exogenous natural output time series Y_exo_true as model target (label). M_exo can adopt a model structure suitable for time series prediction, such as a combination structure of convolutional neural network CNN and recurrent neural network RNN.

[0094] After the model M_exo is trained, it is applied to all meteorological element time series X_met in the multi-dimensional time series input sample set D_clean for inference, thereby generating the exogenous natural output prediction time series Y_exo_hat covering the entire time period.

[0095] Step 203, construct the control effect time series, and build a control effect prediction model to generate a control effect prediction time series.

[0096] The dispatch control effect branch is modeled, specifically including:

[0097] As a preferred embodiment, the specific implementation of this step includes:

[0098] The control effect time series Y_ctrl_true is constructed. Specifically, the historical multi-energy output time series Y_hist_multi is extracted from the multi-dimensional time series input sample set, and the historical multi-energy output time series is differentiated with the exogenous natural output prediction time series Y_exo_hat obtained in the previous step. That is: Y_ctrl_true(t) = Y_hist_multi(t) - Y_exo_hat(t). This Y_ctrl_true represents the control effect actually occurring in history, including scheduling behavior and model error.

[0099] The control effect input feature sequence X_ctrl_feat is constructed. Specifically, the unit and reservoir state time series X_state and the scheduling and control action time series U_ctrl are extracted from the multi-dimensional time series input sample set D_clean, and they are combined as the input features of the model.

[0100] The control effect prediction model M_ctrl is trained. The control effect input feature sequence X_ctrl_feat is used as the model input, and the control effect time series Y_ctrl_true is used as the model target (label) to train the model M_ctrl. The M_ctrl can adopt a gated recurrent neural network GRU or an attention-enhanced sequence model. The time lag selection and feature encoding process of the control effect prediction model can be implemented by using existing technologies, and will not be described in detail here.

[0101] After the model M_ctrl is trained, it is applied to the control effect input feature sequence X_ctrl_feat of the whole sample for inference to generate the control effect prediction time series Y_ctrl_hat.

[0102] In step 204, the exogenous residual time series is calculated based on the difference between the exogenous natural output time series and the exogenous natural output prediction time series.

[0103] This step is used to calculate the uncontrollable exogenous part error. Specifically, the difference between the exogenous natural output time series Y_exo_true (only existing in the D_base period) and the exogenous natural output prediction time series Y_exo_hat is calculated, and the formula is: E_exo(t) = Y_exo_true(t) - Y_exo_hat(t).

[0104] It should be particularly noted that the exogenous residual time series E_exo is only calculated and defined in the time interval corresponding to the baseline control sample set D_base. The E_exo will be the core input of the conformal calibration in Example 3.

[0105] Step 205, calculate the control residual time series based on the difference between the control effect time series and the control effect prediction time series.

[0106] The controllable control part modeling error is calculated, specifically: the difference between the control effect time series Y_ctrl_true and the control effect prediction time series Y_ctrl_hat is calculated, formula: E_ctrl(t)=Y_ctrl_true(t)-Y_ctrl_hat(t).

[0107] Unlike E_exo, the control residual time series E_ctrl is calculated in the full time range. This E_ctrl will not participate in the conformal calibration, but as one of the inputs for the comprehensive credibility evaluation in Embodiment 5.

[0108] Step 206, wherein the exogenous uncertainty to be calibrated data contains the exogenous residual time series and the exogenous natural output prediction time series, and the control effect to be evaluated data contains the control residual time series and the control effect prediction time series.

[0109] This step is to package the data output by the previous steps. The exogenous uncertainty to be calibrated data E_exo is passed to Embodiment 3 for conformal calibration; the control effect to be evaluated data E_ctrl and Y_ctrl_hat are passed to Embodiment 4 and Embodiment 5 for trajectory generation and credibility evaluation.

[0110] Embodiment 3, describes a multi-time scale conformal calibration process based on exogenous residuals.

[0111] As mentioned earlier, this embodiment only calibrates the exogenous residual time series E_exo obtained in Embodiment 2.

[0112] Step 301, determine a time scale set that is suitable for scheduling demand.

[0113] The purpose of this step is to enable the results of conformal calibration to match the actual needs of scheduling business. The time scale set T_set can be determined based on factors such as scheduling rolling period, reserve planning period, and electricity price settlement period. For example, in a system with a fifteen-minute basis sampling step, the time scale set T_set can include: fifteen minutes (for real-time control), one hour (for rolling scheduling), four hours (for climbing and reserve planning), and intra-day peak sensitive period (for critical risk control).

[0114] Step 302, constructing calibration samples on each time scale in the time scale set using the exogenous residual time series in the exogenous uncertainty-to-be-calibrated data, and calculating non-conformal scores to obtain a multi-time scale non-conformal score set.

[0115] This step first needs to aggregate the original exogenous residual time series E_exo. Specifically, for each time scale τ in the time scale set T_set, the E_exo is divided into a sliding window and aggregated according to the time scale (for example, 1 hour), for example, taking the average or maximum value in the window, to form the exogenous residual calibration sample E_exo(τ, i) under the time scale, where i is the window index.

[0116] On this basis, the non-conformal score is calculated. The non-conformal score is an index for quantifying the degree of inconsistency between the predicted value and the true value. In the present embodiment, preferably, the non-conformal score S_exo(τ, i) is defined as the absolute value of the exogenous residual calibration sample under the time scale, that is: S_exo(τ, i) = |E_exo(τ, i)|.

[0117] By traversing all calibration windows i on each time scale τ, the non-conformal score sequence under the time scale is obtained, and the score sequences of all time scales are collected to form the multi-time scale non-conformal score set S_exo_multi.

[0118] Step 303, based on the multi-time scale non-conformal score set, the quantile correction amount of each time scale is calculated under a given confidence level to generate a multi-time scale quantile correction amount set.

[0119] This step is the core of conformal prediction, which is used to determine the correction amount required to meet the target coverage. First, a target confidence level 1-α is set, for example, if α=0.1, it corresponds to a confidence level of 90%.

[0120] For each time scale τ in the multi-time scale non-conformal score set S_exo_multi, the non-conformal score sequence S_exo(τ, i) corresponding thereto is sorted in ascending order. Then, the (1-α) quantile of the sequence is taken as the quantile correction amount Q_exo(τ, α) under the time scale. For example, if there are N scores on a time scale, the quantile correction amount can be taken as the score value after sorting the first upward integer ((N+1)*(1-α)) scores.

[0121] All time scales are traversed to obtain the multi-time scale quantile correction amount set Q_exo_multi.

[0122] Step 304, fusing the multi-time scale quantile correction amount set to generate an exogenous interval correction amount sequence.

[0123] This step aims to fuse the correction amounts at different scales into a unified correction sequence. As a preferred implementation, this fusion process is achieved based on an optimizable time-scale fusion weight w_τ(τ).

[0124] Specifically, first, set the time-scale fusion weight w_τ(τ), where the sum of all weights is one. To determine the optimal weights, a performance metric function can be constructed on the validation set. This function preferably considers both the coverage deviation (i.e., the gap between the actual coverage and the target confidence level 1-α) and the average interval width (the narrower the interval width, the better). By optimizing this performance metric function, such as using grid search or a simple optimization algorithm, a set of optimal time-scale fusion weights w_τ(τ) is determined.

[0125] In this embodiment, the performance metric function can be L(w)=λ cov ∣CR val (w)−(1−α)∣+λ width ⋅W* val (w);

[0126] L(w) is the value of the performance metric function, representing the objective function to be minimized. The smaller the value, the better the fusion effect;

[0127] w is the time-scale fusion weight vector, w = [w _τ1 , w _τ2 , ..., w _τK ; λ cov is the coverage deviation penalty coefficient, used to control the weight of the coverage deviation term, and the recommended value range is from 1 to 10; λ width is the interval width penalty coefficient, used to control the weight of the interval width term, and the recommended value range is from 0.1 to 1; CR val (w) is the actual coverage on the validation set, representing the proportion of the true value falling into the prediction interval under the given weight w; (1-α) is the target confidence level, for example, when α = 0.1, the target confidence level is 90%; W* val (w) is the average normalized interval width on the validation set;

[0128] where, ;

[0129] N_val is the number of samples in the validation set; I_i is the coverage indicator value of the i-th sample, and I_i = 1 when the true value falls into the prediction interval, otherwise I_i = 0.

[0130] ;

[0131] N_val is the number of validation set samples; Δ_exo,i(w) is the fusion interval correction of the i-th sample in the validation set; P_rated is the system rated output;

[0132] After determining the weights, the time-scale fusion weights are applied to the weighted combination of the multi-time-scale quantile correction set to generate the exogenous interval correction sequence Δ_exo(t).

[0133] ;

[0134] Optionally, in order to avoid the adverse effects of the dramatic jump of Δ_exo(t) in time on scheduling, the generated exogenous interval correction sequence Δ_exo(t) can be subjected to time smoothing processing, such as moving average or low-pass filtering.

[0135] Step 305, combining the exogenous interval correction sequence with the exogenous natural output prediction time sequence in the exogenous uncertainty to be calibrated, to construct an exogenous natural output interval prediction result.

[0136] This step uses the correction Δ_exo(t) obtained in the previous step and the exogenous natural output prediction time sequence Y_exo_hat(t) obtained in Example 2 to construct the final exogenous interval. Specifically, the exogenous natural output prediction time sequence Y_exo_hat(t) is taken as the interval center, and the exogenous interval correction sequence Δ_exo(t) is taken as the half-width, to construct the exogenous natural output interval prediction result Y_exo_interval(t):

[0137] Y_exo_interval(t) = [Y_exo_hat(t) - Δ_exo(t), Y_exo_hat(t) + Δ_exo(t)].

[0138] Step 306, wherein the exogenous uncertainty calibration result comprises the exogenous interval correction sequence and the exogenous natural output interval prediction result.

[0139] Y_exo_interval(t) and Δ_exo(t) in this step will be used as the exogenous uncertainty calibration result, which will be passed to Example 4 and Example 5.

[0140] In order to more clearly illustrate steps 303 and 304, a numerical example is provided. It is assumed that the target confidence level is 90% (α = 0.1), and the time scale set T_set contains two scales: τ1 = 15 minutes and τ2 = 1 hour.

[0141] Assume that in step 303, through the calculation of S_exo_multi, the quantile correction quantity Q_exo(τ1, 0.1) = 20 MW of 15-minute scale is obtained; and the quantile correction quantity Q_exo(τ2, 0.1) = 50 MW of 1-hour scale is obtained.

[0142] In step 304, assume that through the optimization of the performance index function on the validation set, the time scale fusion weight determined is: w_τ(τ1) = 0.4, w_τ(τ2) = 0.6.

[0143] Therefore, the generated exogenous interval correction quantity Δ_exo(t) is:

[0144] Δ_exo = (0.4 * 20 MW) + (0.6 * 50 MW) = 8 MW + 30 MW = 38 MW. For the sake of simplicity, it is assumed to be constant on t here,

[0145] If in step 305, the exogenous natural power prediction time series Y_exo_hat(t) = 500 MW at a certain time t, then the finally constructed exogenous natural power interval prediction result Y_exo_interval(t) is: [500 MW - 38 MW, 500 MW + 38 MW], that is, [462 MW, 538 MW].

[0146] In some embodiments, assume that the time scale set contains three time scales, τ1 equal to 15 minutes, τ2 equal to 1 hour, and τ3 equal to 4 hours, that is, K is equal to 3. The target confidence level is set to 90%, that is, α is equal to 0.1. Through historical data statistics, the quantile correction quantities corresponding to each time scale are respectively: Q_exo(τ1, 0.1) is equal to 20 MW, Q_exo(τ2, 0.1) is equal to 45 MW, and Q_exo(τ3, 0.1) is equal to 80 MW. The penalty coefficient is set to λ_cov equal to 5 and λ_width equal to 0.5. The system rated power P_rated is equal to 1000 MW. The number of validation set samples N_val is equal to 100.

[0147] Taking weight scheme B as an example for calculation, the weight vector is set to w_τ1 equal to 0.4, w_τ2 equal to 0.4, and w_τ3 equal to 0.2.

[0148] First, calculate the fusion interval correction quantity. According to the fusion interval correction quantity calculation formula, multiply the weight of each time scale by the corresponding quantile correction quantity and sum them up: Δ_exo is equal to 42 MW.

[0149] Secondly, the actual coverage rate on the validation set is counted. Among the 100 samples of the validation set, the number of samples whose true values fall into the prediction interval is counted. Assuming that the true values of exogenous natural output of 91 samples are counted to fall into the corresponding prediction interval, the actual coverage rate CR_val is equal to 0.91.

[0150] Thirdly, the average normalized interval width on the validation set is calculated. According to the average normalized interval width calculation formula, the interval full width is normalized by dividing by the system rated output: W_val 0.084.

[0151] Finally, the performance index function value is calculated. The calculation results are substituted into the performance index function expression: L(w) is equal to 0.092.

[0152] Using the same calculation method, other candidate weight schemes can be evaluated. For example, the weight scheme A is set as w_τ1 equal to 0.6, w_τ2 equal to 0.3, w_τ3 equal to 0.1, and the calculated fusion interval correction amount Δ_exo is equal to 33.5MW, the actual coverage rate CR_val is equal to 0.85, the average normalized interval width W_val is equal to 0.067, and the performance index function value L(w) is equal to 0.284. The weight scheme C is set as w_τ1 equal to 0.2, w_τ2 equal to 0.5, w_τ3 equal to 0.3, and the calculated fusion interval correction amount Δ_exo is equal to 50.5MW, the actual coverage rate CR_val is equal to 0.94, the average normalized interval width W_val is equal to 0.101, and the performance index function value L(w) is equal to 0.251.

[0153] By comparing the performance index function values of the three weight schemes, it can be seen that the performance index function value of the weight scheme B is the smallest, which is 0.092, so the weight configuration of the weight scheme B is selected as the optimal fusion weight. From the physical meaning, the actual coverage rate of the weight scheme A is 85%, which is lower than the target confidence level 90%, indicating that although the prediction interval is narrow, the statistical effectiveness is insufficient. The actual coverage rate of the weight scheme C is 94%, which is higher than the target confidence level 90%, indicating that the prediction interval is too wide, resulting in a decrease in prediction accuracy. The actual coverage rate of the weight scheme B is 91%, which is close to the target confidence level 90%, and the interval width is moderate, achieving a good balance between statistical effectiveness and prediction accuracy.

[0154] Embodiment 4, describes a multi-energy trajectory generation and screening process embedded with physical constraints.

[0155] This embodiment combines the exogenous uncertainty interval obtained in Embodiment 3 at the statistical level with the physical feasibility constraints of the power grid, generates a set of candidate trajectories that are both statistically reliable and physically feasible through trajectory generation and screening, and lays a foundation for the reliability evaluation in Embodiment 5.

[0156] Step 401, combine the exogenous natural output interval prediction results in the exogenous uncertainty calibration results with the control effect prediction time series in the control effect to be evaluated data, to generate a candidate multi-energy output trajectory set.

[0157] This step is the process of combining exogenous uncertainty with artificial control effect to construct specific output curves. As a preferred embodiment, the specific implementation of this step includes:

[0158] First, construct a candidate control strategy set. The candidate control strategy set is a high-level concept, which can specifically include typical strategies in the dispatch system for simulating different future operation possibilities. For example, this set can include:

[0159] Strong power curtailment strategy, such as setting the upper limit of wind and light curtailment rate to 20%;

[0160] Standard power curtailment strategy, such as setting the upper limit of wind and light curtailment rate to 5%;

[0161] No power curtailment strategy, such as 0% wind curtailment;

[0162] And strong water storage strategy, balanced power generation strategy or strong water release strategy for hydropower, etc.

[0163] Extract a representative exogenous trajectory set from the exogenous natural output interval prediction results Y_exo_interval (for example, [462MW, 538MW] in the example of embodiment 3). Since Y_exo_interval is an interval range, it needs to be sampled in order to be specific for power grid model calculation. For example, the time series corresponding to the upper boundary 538MW, the median 500MW and the lower boundary 462MW of the interval can be extracted as three representative exogenous trajectories. In some more detailed embodiments, Latin hypercube sampling or Monte Carlo sampling methods can also be used to extract N (for example, N=100) representative exogenous trajectories within the interval.

[0164] For any representative exogenous trajectory in the representative exogenous trajectory set, and combined with any candidate control strategy in the candidate control strategy set, adjust the control effect prediction time series Y_ctrl_hat to obtain a control effect adjustment time series. Add it point by point with the representative exogenous trajectory at the corresponding time index to construct a candidate multi-energy output trajectory.

[0165] An example is provided: assume a representative exogenous trajectory is 500 MW at time t, and the control effect prediction Y_ctrl_hat from embodiment 2 is -30 MW, indicating a slight load shedding based on historical data prediction. If the candidate control policy currently adopted is a strong load shedding policy, the control effect adjustment time series at time t can be revised to -100 MW. The value of the resulting candidate multi-energy output trajectory at time t is then 500 MW + (-100 MW) = 400 MW. By traversing all combinations of representative exogenous trajectories and candidate control policies, the set of candidate multi-energy output trajectories is formed, which is also referred to as Y_multi_traj_raw in subsequent steps.

[0166] Step 402, inject the set of candidate multi-energy output trajectories into the physical constraint and operating boundary model, evaluate the degree of violation of physical constraints of the set of candidate multi-energy output trajectories, and obtain a set of physical constraint violation degree indicators.

[0167] This step is the core of the implementation of the physical constraint embedding of the present application. The physical constraint and operating boundary model M_grid is pre-constructed and stored before this step is executed. This model is a high-level concept, and its specific content can include but is not limited to: grid topology data, line impedance parameters, linearized power flow constraints (such as DC power flow (DC-OPF) model or more precise linearized alternating current (AC-OPF) model), unit ramping constraints (such as P_wind(t)-P_wind(t-1)≤Ramp_limit_wind), system reserve constraints (such as ΣP_spare(t)≥Spare_req(t)), and power upper and lower bound constraints of each unit.

[0168] Specifically, for each trajectory in the set of candidate multi-energy output trajectories (which is a time series containing the output values of wind, light, and water over the prediction period 1...T), it is substituted into the M_grid model as a set of known power injections.

[0169] After injection, the model will perform power flow calculation and operating constraint check for each time index t, and count the constraint violation of the trajectory over the entire period. For example, count the number of times or total amount of voltage out-of-limit (V_t>V_max or V_t<V_min), the amplitude and duration of line overload (P_line(t)>P_line_max), the duration of reserve deficiency (Spare(t)<Spare_req), etc. These detailed violation quantities constitute the set of detailed physical constraint violation degree indicators V_phys_detail.

[0170] To facilitate subsequent comparison and selection, these detailed indicators need to be aggregated to form the physical constraint violation index set V_phys. The aggregation method can be varied. For example, different types of violations (power flow, ramping, reserve) can be normalized and weighted to obtain a comprehensive violation score; or, simply, the most severe violation over the entire period (e.g., the maximum line overload percentage) can be taken as the comprehensive physical constraint violation index for that trajectory.

[0171] Step 403: Remove trajectories from the candidate multi-energy output trajectory set whose physical constraint violation exceeds a preset threshold, and form a set of physically feasible trajectories.

[0172] This step is based on the evaluation results of step 402 for screening. The preset threshold is set according to the safety margin and risk preference of the scheduling system. For example, a scheduling system with a very conservative operating strategy may set a zero violation threshold, that is, V_phys must be strictly equal to 0; while a relatively conventional system may set a smaller tolerance threshold, such as allowing short-term (e.g., within 10 minutes) slight (e.g., 105% rated capacity) overloads on the lines.

[0173] Candidate multi-energy output trajectories whose comprehensive physical constraint violation index V_phys exceeds the preset threshold are considered physically infeasible trajectories and removed from the set. The remaining trajectories constitute the set of physically feasible trajectories Y_multi_traj, which will be used for subsequent credibility assessment.

[0174] In some alternative implementations, this step may not employ a hard rejection method. Instead, all candidate trajectories are retained, but their physical constraint violation index V_phys is passed as a penalty or risk factor to subsequent embodiment 5 for comprehensive credibility evaluation. However, in this preferred embodiment, using a rejection method can effectively reduce the computational complexity of subsequent steps.

[0175] Step 404: Based on the set of physically feasible trajectories, the set of physical constraint violation indices, the sequence of exogenous interval corrections in the exogenous uncertainty calibration results, the time series of control residuals in the control effect evaluation data, and the scheduling decision requirements, the multi-energy interval forecast results are generated.

[0176] This step takes the outputs of this embodiment, including Y_multi_traj and V_phys, together with other key outputs (i.e. Δ_exo and E_ctrl), as inputs to perform the credibility assessment and final interval generation. The detailed implementation of this step, including how to utilize these inputs to perform scenario clustering, construct the credibility scoring function, and how to combine with the scheduling requirement D_sched_req to generate the final interval, will be elaborated in Embodiment 5.

[0177] Embodiment 5, describes the decision-aware credibility assessment and interval forecast publishing. That is, the specific process of how to generate the final multi-energy interval forecast result from the set of physically feasible trajectories Y_multi_traj. This embodiment is not simply a statistical output, but introduces the operational scenario clustering and decision-aware credibility assessment, so that the final interval forecast result is more in line with the actual risk preference of the scheduling.

[0178] Step 501, based on the set of physically feasible trajectories, the set of physical constraint violation indicators, the sequence of exogenous interval correction amounts and the time series of control residuals, extract trajectory features.

[0179] This step is to perform feature engineering on the set of physically feasible trajectories Y_multi_traj formed in Embodiment 4, so as to perform clustering and assessment subsequently. The trajectory features are high-level concepts, which can specifically include: the average output level of the trajectory in the key period (such as the midday photovoltaic high period, the evening peak load period), the volatility of the trajectory (such as the maximum climbing rate or the standard deviation), the physical constraint violation indicator V_phys corresponding to the trajectory (which may not be zero even if it is below the threshold), the exogenous interval correction amount Δ_exo corresponding to the trajectory (from Embodiment 3, representing the size of the exogenous uncertainty where the trajectory is located), and the statistics of the control residual time series E_ctrl corresponding to the trajectory (from Embodiment 2, representing the size of the uncertainty of the control model relied on by the trajectory). These features together constitute the set of trajectory features Feature_traj.

[0180] Step 502, cluster the trajectory features to form a set of operational scenario clusters.

[0181] The purpose of this step is to classify operational trajectories with similar features into a category for scenario-based assessment. Conventional clustering algorithms in the art can be used, such as K-Means clustering, DBSCAN density clustering or hierarchical clustering algorithm, to cluster Feature_traj obtained in step 501.

[0182] The formed Cluster_set of operating scenario clusters is a high-level concept, and the clustering result is data-driven. For example, the clustering result can automatically form several clusters, which can correspond to scenarios with clear physical meanings, such as: high wind power output-low physical risk scenario cluster, high wind power output-high flow limit risk scenario cluster, low output-backup tight scenario cluster, or severe water regulation scenario cluster, etc.

[0183] Step 503, constructing an integrated confidence score function, which is used to integrate the quantification of the physical risk derived from the set of physical constraint violation indicators, the exogenous uncertainty derived from the sequence of exogenous interval corrections, and the control modeling error derived from the sequence of control residuals.

[0184] The integrated confidence score function R_conf aims to give a comprehensive score to each trajectory in the set of physically feasible trajectories, and the lower (or higher, depending on the definition) the score represents the higher the confidence.

[0185] As a preferred embodiment, the construction of the score function is adaptive and decision-aware. Specifically, it is necessary to first evaluate the dispatch sensitivity of each operating scenario cluster in the Cluster_set of operating scenario clusters. The dispatch sensitivity refers to the degree of negative impact on system safety or operating cost if the operating state under the scenario cluster deviates. For example, the dispatch sensitivity of the high flow limit risk scenario cluster is significantly higher than that of the low physical risk scenario cluster.

[0186] According to the evaluated dispatch sensitivity, the penalty weights for quantifying the physical risk V_phys, the exogenous uncertainty Δ_exo, and the control modeling error E_ctrl in the integrated confidence score function are adaptively adjusted. In some scenarios, the dispatch sensitivity is the incremental system operating cost caused by unit output deviation under the scenario cluster.

[0187] For example, the function can be designed as a weighted sum:

[0188] R_conf=W_phys*V_phys+W_exo*Δ_exo+W_ctrl*E_ctrl.

[0189] Where W_phys, W_exo, and W_ctrl are penalty weights.

[0190] Preferably, for operation scenario clusters with high dispatch sensitivity, such as high risk of flow violation clusters, the penalty weights of the corresponding physical risk and control modeling error are increased, i.e. the values of W_phys and W_ctrl are increased. For scenario clusters with low dispatch sensitivity, a set of smaller benchmark weights can be used. In this way, the adverse factors in high-risk scenarios are significantly amplified in the score, thereby achieving a decision-aware risk assessment.

[0191] Step 504: Apply the comprehensive confidence score function to the set of physically feasible trajectories to generate a set of source-side comprehensive confidence assessment indicators.

[0192] This step is to perform a calculation. That is, the score function R_conf constructed in step 503, which can vary in clusters, is applied to each trajectory in the set of physically feasible trajectories Y_multi_traj obtained in embodiment 4, and a specific confidence score is calculated. All these scores together constitute the set of source-side comprehensive confidence assessment indicators.

[0193] Step 505: Generate the multi-energy interval forecasting result according to the set of source-side comprehensive confidence assessment indicators and the set of dispatch decision requirement descriptions (referred to as dispatch decision requirements).

[0194] As a preferred implementation, the specific implementation of this step includes:

[0195] Build a comprehensive dispatch evaluation function. This function further integrates the set of dispatch decision requirement descriptions D_sched_req on the basis of the confidence score R_conf (representing safety) in step 504. The D_sched_req includes, for example, economic objectives such as generation cost, curtailment rate, etc. The comprehensive dispatch evaluation function Score_sched can be designed as a weighted sum of R_conf and economic cost (e.g. total generation cost Cost corresponding to the trajectory), for example Score_sched=W_safe*R_conf+W_econ*Cost, for balancing between safety and economy.

[0196] Apply the comprehensive dispatch evaluation function to the set of physically feasible trajectories to generate a final dispatch evaluation score for each trajectory.

[0197] According to the dispatch evaluation score, filter a set of main scenario multi-energy output trajectory Y_multi_traj_main from the set of physically feasible trajectories. For example, the K trajectories with the lowest scores (K can be 3, 5 or 10) can be selected as the main scenarios according to the order from low to high of the dispatch evaluation score (assuming that the lower the score, the better).

[0198] The boundaries of the main scenario multi-energy output trajectory set at each time index are counted to construct the multi-energy interval prediction result Y_multi_interval. For example, at t=1, if the values of the K=3 main scenario trajectories are 80MW, 95MW, and 110MW respectively, the interval boundaries (i.e. the minimum value and the maximum value) obtained by counting are [80MW, 110MW]. At t=2, if the values are 70MW, 80MW, and 90MW respectively, the interval is [70MW, 90MW]. In this way, the interval of the entire prediction period is constructed.

[0199] In some optional embodiments, when constructing the interval, the minimum value and the maximum value (Min / Max) can also not be taken, but the 10th percentile and the 90th percentile can be taken to avoid the interval being excessively widened by individual extreme main scenario trajectories, so as to ensure the safety while taking into account the compactness of the interval.

[0200] The system can also publish the source-side comprehensive confidence evaluation index set R_conf obtained in step 504 and the information of the operation scenario cluster Cluster_set obtained in step 502 together as a scheduling confidence labeling result C_sched, together with the multi-energy interval prediction result Y_multi_interval, to the scheduling system for use.

[0201] Embodiment 6 describes a trajectory-level conformal calibration method based on physical constraint violation degree.

[0202] This embodiment provides another alternative method for generating a physically feasible trajectory set, which is different from Embodiment 3 and Embodiment 4. Instead of regarding conformal calibration and physical constraint screening as two separate steps, it changes the metric of conformal calibration and directly takes the physical constraint violation degree as a non-conformal score, thereby realizing deep integration of statistics and physics.

[0203] In this embodiment, an initial, large-scale candidate multi-energy output trajectory set needs to be generated first. The generation of the set can be similar to the description of step 401 in Embodiment 4, i.e. generated by combining the sampling of exogenous uncertainty and different control strategies, but no strict physical elimination is performed at this time.

[0204] Define the trajectory-level non-conformal score. Unlike Embodiment 3 in which the non-conformal score is defined as the absolute value of the exogenous residual |E_exo(τ,i)|, in this embodiment, the non-conformal score S is defined as the physical constraint violation degree of the entire candidate trajectory Y_1:T.

[0205] Preferably, the score S(Y_1:T) can be represented by a weighted sum formula, for example: S(Y_1:T)=Σ[t=1,T](λ1·vio _tpf + λ2 · vio _t ramp + λ3 · vio _t reserve ).

[0206] where Y_1:T denotes the whole trajectory from time 1 to T; vio _t pf is the voltage or line flow over-limit amount at time t (e.g. the positive part of (P_line(t) - P_line_max)); vio _t ramp is the ramping constraint violation amount at time t; vio _t reserve is the reserve deficiency amount at time t; λ1, λ2, λ3 are weight coefficients for balancing the importance of different types of constraints.

[0207] Construct a trajectory-level conforming prediction set. With the calibration set on historical data (i.e. a set of historical trajectories and their corresponding non-conforming scores S), calculate the (1-α) quantile of these S scores at a given confidence level α (e.g. α = 0.1), obtaining the quantile correction amount q_α.

[0208] Accordingly, the constructed conforming prediction set C_α is no longer a simple upper and lower bound (such as Y_exo_interval in embodiment 3), but a set of trajectories. Its definition is: C_α = {Y_1:T: S(Y_1:T) ≤ q_α}.

[0209] The conforming prediction set C_α has clear physical and statistical significance: it contains all trajectories whose physical constraint violation degree is lower than the threshold q_α, and this set can statistically cover the future real and physically feasible trajectories with a probability of (1-α).

[0210] The risk function is unified with the non-conforming score. Based on this embodiment, when performing credibility evaluation in the subsequent steps, the risk function R(Y_1:T) used can be directly constructed based on (or equivalent to) the non-conforming score S(Y_1:T) in this embodiment, achieving the unification of evaluation indexes.

[0211] In summary, this embodiment uses the physical constraint violation degree as the metric for conforming calibration, so that the generated prediction set naturally tends to filter out trajectories with severe physical violations. The output C_α set can directly replace the physically feasible trajectory set output in step 403 of embodiment 4, and be used for subsequent evaluation in embodiment 5.

[0212] Embodiment 7 describes a scenario cluster adaptive conforming calibration method for decision-making.

[0213] The procedure of Example 3 and Example 5 is to first perform a uniform conformal calibration, and then perform clustering and evaluation on the generated trajectories. This embodiment restructures the procedure, adopting a strategy of first clustering, and then performing adaptive conformal calibration on different clusters.

[0214] Step 701, scene representation and pre-clustering. This step is performed before conformal calibration.

[0215] Construct a scene representation network, which is preferably a deep neural network model, such as an autoencoder or a graph neural network.

[0216] The input of the network is the system state feature at the current time, which is a high-level concept, and can specifically include:

[0217] Meteorological features, such as wind speed, irradiance;

[0218] Power grid state, such as key line load rate;

[0219] Load level, and possibly market price signals, etc. The network embeds these high-dimensional inputs into a low-dimensional scene representation z space.

[0220] In the z space, a clustering algorithm such as K-Means is used to cluster historical samples to form several scene clusters. Unlike the clustering in Example 5, the clusters here are divided based on input features, rather than based on output trajectories. For example, the clustering results can form a high-wind-high-load-grid-easy cluster (cluster k = 1), a low-wind-high-load-grid-congested cluster (cluster k = 2), etc.

[0221] Step 702, cluster-specific conformal calibration. This step is the core of this embodiment. In Example 3, the quantile correction quantity Q_exo(τ,α) of the non-conformal score is calculated on all calibration samples. But in this embodiment, conformal calibration is performed on a cluster-by-cluster basis.

[0222] Specifically, for each scene cluster k formed in step 701, only the calibration samples belonging to the cluster k are used to calculate its corresponding non-conformal score sequence {S(k,i)}, and the score S can be |E_exo| in Example 3, or S(Y_1:T) in Example 6, and its quantile correction quantity q_k(α) is calculated separately.

[0223] For example, in the grid-congested cluster k = 2, its historical residual or physical violation degree may be naturally higher, so its correction quantity q_k=2(α) will be significantly greater than the correction quantity q_k=1(α) of the grid-easy cluster k = 1.

[0224] Step 703, decision-aware parameter adjustment. Preferably, the conformal calibration parameters of the cluster k can be further adjusted adaptively according to the scheduling sensitivity of the cluster k. For example, for the congested cluster k = 2 (high scheduling sensitivity), the system can automatically increase its confidence level, for example, from 90% to 95%, that is, a from 0.1 to 0.05, or give higher weight λ to the physical violation item when calculating its non-conformal score (such as S in Embodiment 6).

[0225] In actual prediction, the system first determines which scene cluster k the current belongs to according to the input features at the current time, and then directly calls the pre-computed quantile correction q_k(a) or the adjusted parameters corresponding to the cluster k to construct the prediction interval.

[0226] It should be noted that in the present application, some models, such as prediction models, fusion models, deep neural network models, etc., the specific type, loss function, optimization method and hyperparameters, etc. Technical details can be implemented by existing methods, and the present application will not be described in detail. In addition, the confidence level, the fusion weight, the physical constraint threshold, the clustering parameter, etc. Can be obtained by using existing methods by those skilled in the art, and the numerical value is different in different scenarios, and will not be described here.

[0227] In summary, by prepositioning the scene clustering, the embodiment realizes the adaptation and differentiation of the conformal calibration parameters, so that the prediction interval can dynamically adjust its width according to the current system running state, such as congestion or high risk, to realize the truly decision-aware conformal prediction.

Claims

1. A wind-solar-water multi-energy hybrid system source-side key time series interval forecasting and credibility evaluation method, characterized in that, include: Obtain a multi-dimensional time-series input sample set; Based on a multi-dimensional time-series input sample set, a causal structure model is constructed to decouple exogenous natural output and scheduling control effects, thereby obtaining exogenous uncertainty data to be calibrated and control effect data to be evaluated. Perform multi-timescale conformal calibration only on the data to be calibrated for exogenous uncertainties to obtain the calibration results for exogenous uncertainties. By integrating the calibration results of exogenous uncertainties, the data to be evaluated on the control effects, the pre-stored physical constraints and operational boundary models, and the scheduling decision requirements, multi-energy range forecast results are generated. A causal structure model is constructed to decouple exogenous natural output from scheduling control effects, yielding exogenous uncertainty data to be calibrated and control effect data to be evaluated, including: Construct a causal structure model; Based on the causal structure model, an exogenous natural output prediction model is constructed, and exogenous natural output time series and exogenous natural output prediction time series are generated. Construct a time series of control effects and build a control effect prediction model to generate a time series of control effect predictions; Based on the difference between the exogenous natural power output time series and the exogenous natural power output prediction time series, the exogenous residual time series is calculated, and together with the exogenous natural power output prediction time series, they form the exogenous uncertainty data to be calibrated. Based on the difference between the control effect time series and the control effect prediction time series, the control residual time series is calculated and combined with the control effect prediction time series to form the control effect data to be evaluated. The steps for generating multi-energy range forecast results, which integrate exogenous uncertainty calibration results, control effect data to be evaluated, physical constraints and operational boundary models, and scheduling decision requirements, include: By combining the exogenous natural output range prediction results in the exogenous uncertainty calibration results with the control effect prediction time series in the control effect to be evaluated data, a candidate multi-energy output trajectory set is generated. The candidate multi-energy output trajectory set is injected into the physical constraint and operational boundary model to evaluate the degree of physical constraint violation of the candidate multi-energy output trajectory set, and obtain the physical constraint violation index set. Trajectories in the candidate multi-energy output trajectory set that violate physical constraints to a degree exceeding a preset threshold are removed, forming a set of physically feasible trajectories; Based on the set of physically feasible trajectories, the set of physical constraint violation indices, the sequence of exogenous interval corrections in the exogenous uncertainty calibration results, the time series of control residuals in the data to be evaluated for control effects, and scheduling decision requirements, multi-energy interval forecast results are generated.

2. The method of claim 1, wherein, Construct an exogenous natural power output prediction model and generate exogenous natural power output time series and exogenous natural power output prediction time series, including: Establish a baseline control decision rule, which is used to identify intervals in the multidimensional time-series input sample set that are close to no control intervention; The benchmark control sample set is obtained by filtering the multi-dimensional time series input sample set using the benchmark control judgment rules. The historical multi-energy output time series contained in the specified baseline control sample set are exogenous natural output time series. The exogenous natural output prediction model is trained using only the baseline control sample set.

3. The method of claim 1, wherein, Constructing a control effect time series and building a control effect prediction model to generate a control effect prediction time series, including: extracting a historical multi-energy output time series from the multi-dimensional time-series input sample set, and differentiating the historical multi-energy output time series from the exogenous natural output prediction time series to construct a control effect time series; extracting a unit and reservoir state time series and a scheduling and control action time series from the multi-dimensional time-series input sample set to construct a control effect input feature sequence; training the control effect prediction model by taking the control effect input feature sequence as the model input and the control effect time series as the model target.

4. The method of claim 1, wherein, The steps of constructing the causal structure model include: combining the physical mechanism and operation experience of the wind-solar-hydro multi-energy hybrid system to set an initial causal structure model, and the initial causal structure model explicitly distinguishes between exogenous driving paths and scheduling control paths; based on the multi-dimensional time-series input sample set, a data-driven causal discovery method is used to modify and verify the initial causal structure model to obtain the causal structure model.

5. The method of claim 1, wherein, The steps of performing multi-time scale conformal calibration to obtain the exogenous uncertainty calibration result include: determining a time scale set suitable for scheduling requirements; using the exogenous residual time series in the exogenous uncertainty to be calibrated data, constructing a calibration sample at each time scale in the time scale set, and calculating a non-conformal score to obtain a multi-time scale non-conformal score set; based on the multi-time scale non-conformal score set, the quantile correction amount of each time scale is calculated under a given confidence level to generate a multi-time scale quantile correction amount set; fusing the multi-time scale quantile correction amount set to generate an exogenous interval correction amount sequence; combining the exogenous interval correction amount sequence and the exogenous natural output prediction time series in the exogenous uncertainty to be calibrated data to construct an exogenous natural output interval prediction result; wherein the exogenous uncertainty calibration result comprises the exogenous interval correction amount sequence and the exogenous natural output interval prediction result.

6. The method of claim 5, wherein, The steps of fusing the multi-time scale quantile correction amount set to generate the exogenous interval correction amount sequence include: setting a time scale fusion weight; on the validation set, a performance index function is constructed with coverage bias and interval width as the target; determining the time scale fusion weight by optimizing the performance index function; applying the time scale fusion weight to the multi-time scale quantile correction amount set to generate the exogenous interval correction amount sequence.

7. The method of claim 1, wherein, The steps of generating a candidate multi-energy output trajectory set include: constructing a candidate control strategy set; extracting a representative exogenous trajectory set from the exogenous natural output interval prediction result; for any representative exogenous trajectory in the representative exogenous trajectory set, and combining any candidate control strategy in the candidate control strategy set, adjusting the control effect prediction time series to obtain a control effect adjustment time series; adding the control effect adjustment time series and the representative exogenous trajectory point by point to construct a candidate multi-energy output trajectory, and then forming a candidate multi-energy output trajectory set.

8. The method of claim 7, wherein, The steps of generating a multi-energy interval prediction result based on the physical feasible trajectory set, the physical constraint violation degree index set, the exogenous interval correction amount sequence, the control residual time series, and the scheduling decision requirement description set further include: Based on the set of physically feasible trajectories, the set of physical constraint violation indexes, the sequence of exogenous interval correction amounts and the time series of control residuals, trajectory features are extracted; The trajectory features are clustered to form a set of operating scenario clusters; A comprehensive confidence score function is constructed, which is used to integrate the physical risk derived from the set of physical constraint violation indexes, the exogenous uncertainty derived from the sequence of exogenous interval correction amounts, and the control modeling error derived from the time series of control residuals; The comprehensive confidence score function is applied to the set of physically feasible trajectories to generate a set of source-side comprehensive confidence assessment indexes; And according to the set of source-side comprehensive confidence assessment indexes and the set of scheduling decision requirement descriptions, the multi-energy interval prediction results are generated.

Images