Photovoltaic power prediction method based on hierarchical, admission and predictive continuity control

Abstract

The application discloses a photovoltaic power prediction method based on hierarchical, access and prediction continuity control, which comprises the following steps: S1, collecting information; S2, constructing a prediction condition characteristic vector and calculating a prediction support index; S3, hierarchical evaluation, obtaining a prediction feasibility grading result of a current prediction task; S4, dynamically determining a prediction organization mode suitable for the current prediction task and outputting a prediction mode; S5, constructing an expert access constraint matrix and forming a candidate prediction expert set; S6, selecting a leading prediction expert and an auxiliary prediction expert in the candidate prediction expert set; S7, combining the output results of the leading prediction expert and the auxiliary prediction expert to form an initial prediction result; S8, implementing prediction continuity control on the initial prediction result and a generation path thereof; and S9, combining engineering constraints and hierarchical coordination information to form a final prediction result. The application can improve the stability, continuity and engineering usability of photovoltaic power prediction under complex conditions.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a photovoltaic power prediction method based on hierarchical, access and prediction continuity control. Background Technology

[0002] Photovoltaic power forecasting plays a crucial role in grid dispatching applications, output planning, operation optimization, and electricity market transactions. In practical applications, photovoltaic power plant forecasting tasks are not always conducted under ideal conditions of complete input, sufficient historical data, and stable scenarios. Common challenges include missing key meteorological data, insufficient historical samples for target equipment, interrupted observation of the target object, rapid changes in operating scenarios, and the superposition of multiple adverse conditions. These factors can significantly alter the applicability of the same forecasting model across different tasks and time periods, thereby affecting forecasting accuracy, stability, and engineering usability.

[0003] Most existing photovoltaic power prediction methods are based on fixed input structures and fixed prediction models, typically only performing local adaptations for a single missing factor or employing simple degradation strategies to maintain the prediction output. These methods lack a unified mechanism for representing the feasibility of the current prediction task, and also lack the technical means to select prediction models based on current information support and control the model switching process. Therefore, under conditions of incomplete meteorological data, short historical data, complex disturbances, and rapid scene switching, they are prone to model mismatch, abrupt switching, and unstable prediction results. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides a photovoltaic power prediction method based on hierarchical, admission and prediction continuity control. By combining prediction feasibility classification, prediction mode determination, candidate prediction expert admission and main and auxiliary guarantee expert organization, a prediction path matching the current task can be formed under different information support conditions, so as to improve the stability of photovoltaic power prediction under complex conditions.

[0005] Technical Solution: To achieve the above objectives, the photovoltaic power prediction method based on hierarchical, access, and predictive continuity control of the present invention includes the following steps: S1: Collect prediction input information corresponding to the target power plant or target equipment to form the original input set of the current prediction task; S2: Based on the original input set of the current prediction task, extract feature quantities to characterize the prediction conditions of the current prediction task, construct a prediction condition feature vector, and calculate prediction support indicators based on the prediction condition feature vector; S3: Based on the prediction condition feature vector and prediction support indicators, perform hierarchical evaluation of the feasibility of the current prediction task at the current moment to obtain prediction feasibility classification results; S4: Based on the prediction condition feature vector, prediction support indicators, and prediction feasibility classification results, construct a prediction mode state machine, dynamically determine the prediction organization method suitable for the current prediction task under different conditions and operating conditions, and output the prediction mode at the current moment; S5: Refer to the prediction functional unit as a prediction expert, and define the capability boundary of the prediction expert. Description information: Based on the prediction support indicators, the current prediction feasibility level, the current target prediction mode, and task requirements, an expert admission constraint matrix is ​​constructed to screen prediction experts and form a candidate prediction expert set accordingly; S6: Within the candidate prediction expert set, the expert adaptation score for the current task is calculated for each prediction expert, and the prediction expert with the highest score is selected as the leading prediction expert. From the remaining candidate prediction experts, one or more prediction experts with high scores and complementary functions are selected as auxiliary prediction experts; S7: The leading prediction expert outputs the main prediction result based on the current prediction task, and the auxiliary prediction experts output the compensation or correction amount; Combining the output results of the leading prediction expert and the auxiliary prediction experts, the initial prediction result at the current moment is formed; S8: Prediction continuity control is implemented on the initial prediction result and its generation path; S9: Engineering constraint verification is performed on the prediction result after prediction continuity control, and consistency correction is performed in combination with the hierarchical collaboration information of the current prediction task to form the final prediction result.

[0006] Furthermore, in S2, time... Predictive conditional feature vector Represented as: ;in, Indicates key observation integrity characteristics, This indicates the sufficiency of historical support. This indicates the availability characteristics of auxiliary support. This indicates the characteristics of the operating condition disturbance. This indicates hierarchical collaborative support characteristics. Represents task constraint characteristics; time step Forecast support indicators Represented as: ;in, Represents the feature mapping function; predicts supporting indicators. Expanded by dimension: ;in, Indicates key observational support indicators, Indicates historical support indicators. This indicates a secondary support indicator. This indicates the operating condition disturbance index. Indicators representing hierarchical collaborative support. This indicates the task constraint indicators.

[0007] Furthermore, in S3, time... Predictive feasibility score Represented as: ;in, This represents the adjustment coefficient for the corresponding dimension. This represents the first result obtained from the predictive support indicators. Support evaluation results This indicates a penalty for scenarios with multiple adverse conditions; based on the prediction feasibility score, the current prediction task is divided into different prediction feasibility levels: ;in, Indicates time The predicted feasibility level, This represents the mapping function from the predicted feasibility score to the grade label.

[0008] Furthermore, in S4, pattern fit is represented as: ;in, Indicates the first Predictive patterns at time Mode compatibility; Indicates the relationship with the first The feature contribution coefficient vector corresponding to the prediction pattern; Indicates time The predicted conditional feature vector; Indicates the first The entry constraint violation of the prediction model under the current conditions is used to characterize the degree of non-compliance of basic conditions such as key observation lower limit, historical sample lower limit, availability of auxiliary support and stability of operating conditions. This represents the corresponding penalty coefficient; the prediction pattern at the current time is represented as: ;in, Indicates time Predictive patterns, This indicates the prediction pattern from the previous moment. Indicates the current feasibility level of the forecast. This represents the current set of adaptability scores for each mode. This represents mode switching hysteresis constraints and boundary buffer constraints. This indicates the duration of the current state. This represents the state transition function.

[0009] Furthermore, in S5, the first A vector describing the capabilities of a prediction expert Represented as: ;in, , , and They represent the first The minimum requirements for key observational support, historical support, auxiliary support, and hierarchical collaborative support for forecasting experts. and These represent the upper limits of operating condition disturbances and task constraints that the forecasting expert can tolerate, respectively. This indicates the applicable forecasting feasibility level, forecasting model, forecasting level, and forecasting baseline range for the forecasting expert; A forecasting expert at a certain moment The admission results are expressed as follows:

[0010] ;

[0011] in, Indicates the first A forecasting expert at a certain moment Possess the necessary qualifications. This indicates that the applicant does not meet the eligibility requirements. This indicates the prediction feasibility level of the current prediction task. This indicates the prediction mode for the current prediction task. This indicates the prediction level of the current prediction task. This represents the prediction timescale or prediction horizon of the current prediction task; therefore, the expert admission constraint matrix at the current moment is expressed as: Based on this, a set of candidate prediction experts is formed: .

[0012] Furthermore, in S6, the expert fit score is expressed as: ;in, Indicates the first A forecasting expert at a certain moment Expert matching score, This indicates the current forecasting task's supporting foundation and the first The degree of matching between the descriptions of the predictive experts' abilities. Indicates the first The evaluation of a predictive expert's local performance in a nearby historical context. This represents the output stability evaluation value of the forecasting expert under the current forecasting model. This indicates a switch to the continuity acceptance cost introduced by the forecasting expert. , , and The adjustment coefficient is used; based on the expert fit score, the predictive expert with the highest score is selected as the lead predictive expert, denoted as:

[0013] .

[0014] Furthermore, in S6, a conservative prediction expert is selected from the set of candidate prediction experts, exhibiting high output stability and strong continuous transition capability. Under normal conditions, the conservative prediction expert does not directly participate in the generation of the main prediction result but remains available for use. When instability of the dominant prediction expert, failure of candidate expert switching, prediction mode degradation, or the current task entering a high-uncertainty scenario is detected, the conservative prediction expert participates in the prediction output as a fallback prediction path to ensure the continuity and basic availability of the prediction results. The conservative prediction expert is represented as follows:

[0015] ;

[0016] in, Indicates the first The ability of a forecasting expert to maintain forecast continuity.

[0017] Furthermore, in S7, the initial prediction result at the current time step is expressed as:

[0018] ;

[0019] in, This indicates the main forecast result output by the leading forecasting expert. This represents a set of experts assisting in prediction. Indicates the first The compensation or correction amount provided by an auxiliary forecasting expert. This represents the synergistic effect coefficient of the corresponding auxiliary prediction expert under the current conditions.

[0020] Furthermore, in S8, when implementing forecast continuity control, it first determines whether there is a forecast continuity risk at the current moment, and then performs parallel verification on the new forecasting expert or new forecasting path to replace the current dominant forecasting path. If the parallel verification meets the preset conditions, the original forecasting output and the new forecasting output are combined according to the continuity control coefficient within the transition window to generate the forecasting result after continuous transition. If the parallel verification does not meet the preset conditions, it reverts to the output result of the original dominant forecasting expert or conservative forecasting expert. In S8, the forecasting result after continuous transition is expressed as follows: ;in, This indicates the output result of the original dominant prediction path. This indicates the output of the new predicted path. This represents the continuous control coefficient, which is adjusted according to a monotonically changing pattern within the transition window; to constrain the magnitude of change in the predicted output during the transition phase, the following must be satisfied: ;in, Indicates time The prediction results This indicates the prediction result at the previous moment. Indicates time The permissible change boundary.

[0021] Furthermore, in S9, the final prediction result is obtained by the following optimized form:

[0022] ;

[0023] in, This indicates the final prediction result. This represents the set of feasible solutions that satisfy the engineering constraints. This represents the predicted outcome variable to be optimized. This indicates the forecast result after continuous forecast control. Represents a hierarchical aggregation mapping matrix. This represents the measured value, reliable value, or reliable predicted value of the upper-layer device or object. This indicates the reference value for the same floor, the structural reference value, or the historical structural allocation reference value. and This is the adjustment coefficient.

[0024] Beneficial Effects: The photovoltaic power prediction method based on hierarchical, admission, and predictive continuity control of the present invention has the following beneficial effects:

[0025] 1) By constructing prediction condition feature vectors and prediction support indicators, the completeness of key observations, the sufficiency of target history, the availability of auxiliary information, the volatility of operating conditions, the hierarchical collaborative support capability and task constraints are uniformly represented. On this basis, the prediction feasibility classification and prediction mode determination are completed, so that the system can adaptively select a more suitable prediction path for different prediction conditions and avoid the mismatch problem of fixed models or fixed processes in complex scenarios.

[0026] 2) Based on the capability boundary of prediction experts, the admission limit is set for candidate prediction experts, and the leading prediction experts, auxiliary prediction experts and conservative prediction experts are determined within the admission range. This enables different prediction experts to undertake the main prediction, compensation correction, boundary verification or conservative output tasks under more suitable prediction conditions, thereby improving the stability and applicability of photovoltaic power prediction in observation-limited, short-history, auxiliary support and boundary scenarios.

[0027] 3) By implementing forecast continuity control, smooth constraints are imposed on changes in forecast feasibility level, forecast model, leading forecast expert, and forecast output transition process. Furthermore, by combining non-negative constraints, capacity limit constraints, nighttime near-zero constraints, sunrise and sunset transition constraints, and credibility assessment output, the risk of jumps and jitters in forecast results at model boundaries and expert replacement boundaries is reduced, thereby improving the engineering availability of forecast results in scheduling applications, operational analysis, and business applications. Attached Figure Description

[0028] Appendix Figure 1 This is a flowchart of the photovoltaic power prediction method of the present invention. Detailed Implementation

[0029] The invention will now be further described with reference to the accompanying drawings.

[0030] As attached Figure 1 The photovoltaic power prediction method based on hierarchical, access, and prediction continuity control, as described herein, clarifies that the "expert" or "prediction expert" referred to is not a human expert with experience, but rather a prediction functional unit designed for specific input dependency conditions, historical support levels, auxiliary support availability, operating condition characteristics, and prediction task requirements. This prediction functional unit can consist of an independent prediction model, a dedicated prediction branch under a shared backbone, a prediction module combining rules and models, an error correction module, a hierarchical collaborative correction module, or other computational units capable of outputting prediction results under corresponding conditions. Different prediction experts have different input dependency boundaries, historical sample requirements, auxiliary support dependency levels, boundary scenario adaptability, output stability, and prediction continuity capabilities. The photovoltaic power prediction method of this invention includes the following steps.

[0031] S1: Standardization of multi-source data access and prediction tasks.

[0032] Collect the prediction input information corresponding to the target power plant or target equipment to form the original input set for the current prediction task. The input information includes at least historical power data, meteorological observation data, equipment static attribute data, time and location parameters, operating status identification information, and optional auxiliary support information.

[0033] The historical power data includes power station-level power, inverter-level power, string-level power, or historical output sequences of other target objects.

[0034] The meteorological observation data includes irradiance, ambient temperature, component temperature, wind speed, wind direction, humidity, cloud cover, or other observations that can characterize environmental conditions.

[0035] The static attribute data of the equipment includes rated capacity, installation orientation, tilt angle, component type, inverter model, equipment category, geographical location, and hierarchical relationship.

[0036] The time and location parameters include time, date, season information, solar altitude angle, and solar azimuth angle; the operating status identification information includes status information such as limited power generation, maintenance, offline, fault, communication abnormality, occlusion abnormality, or other status information affecting the interpretation and result constraints of prediction.

[0037] The auxiliary support information includes reference sequences, aggregate statistics, statistical characteristics of devices at the same level, collaborative information of upper or lower level devices, regional prior information, and other support information that can assist the current prediction task.

[0038] The above input information is processed through time alignment, field mapping, unit unification, abnormal null value identification, duplicate record cleanup, and basic quality checks to form a prediction task object with a unified structure.

[0039] Furthermore, by combining the target prediction level, prediction horizon, time resolution, and prediction output caliber, the current prediction task is standardized to ensure that subsequent prediction feasibility classification, prediction mode determination, expert access judgment, prediction continuity control, level consistency verification, and result output are all carried out under a unified task caliber.

[0040] Furthermore, when key observations, historical sequences, or supporting information are missing, abnormal, or continuously interrupted, supplementary information that has undergone reliable reconstruction, quality labeling, or validity screening can be introduced as auxiliary input to enhance the support foundation for the prediction task; however, the supplementary information does not change the true availability record of the original input information, and subsequent prediction feasibility grading and expert admission judgment are preferentially based on the original availability record.

[0041] S2: Construct the prediction condition feature vector and prediction support index.

[0042] Based on the original input set of the current prediction task, feature quantities characterizing the prediction conditions of the current prediction task are extracted to construct a prediction condition feature vector. This prediction condition feature vector is used for subsequent prediction feasibility classification, prediction mode determination, prediction expert admission, and prediction continuity control. The prediction condition features include at least key observation integrity features, historical support sufficiency features, auxiliary support availability features, operating condition disturbance features, hierarchical collaborative support features, and task constraint features.

[0043] Among them, the key observation integrity features are used to characterize the completeness rate, continuous missing duration, missing type, direct usability and observation reliability of key meteorological quantities, state quantities and target object observations in the current forecasting task.

[0044] Historical support sufficiency features are used to characterize the length of historical samples, the proportion of effective samples, the continuous effective interval, the degree of recent historical stability, and the sufficiency of historical context available for the current prediction task for the target prediction object.

[0045] Operating condition disturbance characteristics are used to characterize the stability, volatility, and boundary disturbance characteristics of the current operating scenario, including nighttime, sunrise and sunset, rapid changes in cloud cover, extreme weather, power outages, faults, or other abnormal operating conditions.

[0046] Auxiliary support availability features are used to characterize the availability, effectiveness, and stability of reference sequences, aggregate statistics, operational status labels, peer reference information, upper and lower layer collaborative information, or other auxiliary support information.

[0047] Hierarchical collaborative support features are used to characterize whether the current target prediction object has upper-level aggregation information, lower-level sub-item information, same-level reference information, or historical structure allocation information that can be used for collaborative correction or consistency verification.

[0048] Task constraint features are used to characterize the implementation constraints formed by the prediction horizon, prediction level, output time scale, granularity, and real-time requirements on the current prediction task.

[0049] time Predictive conditional feature vector Represented as: ;in, Indicates key observation integrity characteristics, This indicates the sufficiency of historical support. This indicates the availability characteristics of auxiliary support. This indicates the characteristics of the operating condition disturbance. This indicates hierarchical collaborative support characteristics. This represents the task constraint characteristics.

[0050] Furthermore, to enhance the stability of prediction feasibility grading, prediction model determination, and prediction expert access control, prediction support indicators are calculated based on prediction condition feature vectors. These prediction support indicators characterize the supporting foundation of the current prediction task across different dimensions, at any given time. Forecast support indicators Represented as: ;in, This represents the feature mapping function. The feature mapping function can be a linear mapping, a piecewise mapping, a regular mapping, a statistical learning mapping, or a combination thereof.

[0051] Furthermore, to facilitate subsequent feasibility grading and expert evaluation for forecasting, forecasting support indicators can be... Expanded by dimension: ;in, Indicates key observational support indicators, Indicates historical support indicators. This indicates a secondary support indicator. This indicates the operating condition disturbance index. Indicators representing hierarchical collaborative support. This indicates the task constraint indicators.

[0052] S3: Predictive feasibility classification.

[0053] Based on the predictive condition feature vector and predictive support indicators, the feasibility of the current predictive task at the current moment is evaluated in a graded manner, resulting in a predictive feasibility grading result. This predictive feasibility grading characterizes the comprehensive feasibility of the current predictive task under the combined effects of key observation integrity, historical support sufficiency, auxiliary support availability, operational condition disturbance degree, hierarchical collaborative support capability, and task constraints. It serves as an important basis for subsequent predictive model determination, predictive expert admission, and predictive continuity control.

[0054] The prediction feasibility classification should take into account at least the following factors: (1) whether the key observations are complete, continuous and reliable; (2) whether the historical samples of the target object are sufficient, continuous and stable; (3) whether the auxiliary support information is available and has effective reference value; (4) whether the current working condition is stable and whether there are rapid fluctuations, boundary disturbances or abnormal operating scenarios; (5) whether there is upper-level aggregation information, lower-level sub-item information or same-level reference information that can be used for hierarchical collaborative correction; (6) whether the prediction level, prediction horizon, time resolution and real-time requirements increase the difficulty of realizing the current prediction task.

[0055] time Predictive feasibility score Represented as: ;in, This represents the adjustment coefficient for the corresponding dimension. This represents the first result obtained from the predictive support indicators. Support evaluation results This indicates penalties for critical missing measurements, insufficient historical data, unstable auxiliary support, abnormal operating conditions, and multiple adverse conditions occurring simultaneously.

[0056] Furthermore, based on the prediction feasibility score, the current prediction task is divided into different prediction feasibility levels: ;in, Indicates time The predicted feasibility level, This represents the mapping function from the predicted feasibility score to the grade label.

[0057] The prediction feasibility levels include at least: (1) Level 1 prediction feasibility, i.e., high feasibility level, which is applicable to prediction tasks with relatively complete key observations, sufficient historical support and relatively stable current operating conditions; (2) Level 2 prediction feasibility, i.e., medium feasibility level, which is applicable to prediction tasks with limited key observations but still with strong historical support; (3) Level 3 prediction feasibility, i.e., limited feasibility level, which is applicable to prediction tasks with a short history of the target object, insufficient effective samples or poor historical continuity; (4) Level 4 prediction feasibility, i.e., auxiliary feasibility level, which is applicable to prediction tasks with insufficient key observations or historical support of the target object itself, but with high availability of support information at the same level, above and below or reference levels; (5) Level 5 prediction feasibility, i.e. conservative feasibility level, which is applicable to prediction tasks with multiple adverse conditions superimposed, high uncertainty at the boundary or extreme operating conditions.

[0058] It should be noted that the prediction feasibility level obtained in this step is used to characterize the feasibility of the current prediction task under the existing support conditions. It focuses on the strength of the current task's supporting foundation and the range of prediction capabilities that can be achieved. The subsequent prediction model determination in S104 is used to determine the prediction path and prediction organization method that should be given priority under this feasibility level. The two are related but not the same.

[0059] S4: Predictive pattern determination based on state machine.

[0060] Based on the prediction condition feature vector, prediction support index, and prediction feasibility classification results, a prediction mode state machine is constructed to dynamically determine the prediction organization method suitable for the current prediction task under different input conditions, historical conditions, auxiliary support conditions, and operating conditions, and output the prediction mode at the current moment.

[0061] The prediction mode state machine includes at least the following states: (1) Complete information state, which is suitable for prediction tasks with relatively complete key observations, sufficient historical samples of the target object, and relatively stable current operating conditions; (2) Limited observation state, which is suitable for prediction tasks with missing key observations but the historical data of the target object can still provide the main prediction support; (3) Short history state, which is suitable for prediction tasks with a short historical window of the target object, insufficient effective samples, newly connected equipment, or poor historical continuity; (4) Auxiliary support state, which is suitable for prediction tasks with insufficient key observations or historical support of the target object itself, but with relatively available auxiliary reference information, aggregated statistical information, hierarchical collaborative information, or operation identification information; (5) Conservative output state, which is suitable for prediction tasks with multiple adverse conditions superimposed, or currently in boundary scenarios such as night, near night, extreme weather, abnormal fluctuations, limited issuance, or failure.

[0062] Unlike static classification methods, the prediction mode state machine sets entry conditions, maintenance conditions, and exit conditions for each state, and introduces hysteresis thresholds, minimum dwell time, and boundary buffers to reduce the risk of frequent state switching near critical conditions and improve the stability of prediction path organization.

[0063] Furthermore, the state transition is not only constrained by the current prediction feasibility level, but also limited by the minimum dwell time of the current mode, the allowed transition relationship between adjacent modes, mode adaptability, and boundary buffer conditions; for mode switching requests that do not meet the transition prerequisites, even if the mode adaptability of the target mode is higher than that of the current mode, the system will still keep the current prediction mode unchanged until the preset transition conditions are met.

[0064] Regarding the first Predictive patterns, time Pattern adaptability can be expressed as: ;in, Indicates the first Predictive patterns at time Mode compatibility; Indicates the relationship with the first The feature contribution coefficient vector corresponding to the prediction pattern. Representing vectors transpose; Indicates time The predicted conditional feature vector; Indicates the first The entry constraint violation of the prediction model under the current conditions is used to characterize the degree of non-compliance of basic conditions such as key observation lower limit, historical sample lower limit, availability of auxiliary support and stability of operating conditions. This represents the corresponding penalty coefficient.

[0065] Furthermore, the prediction model at the current moment is jointly determined by the state at the previous moment, the current prediction feasibility level, the current model fit, and state transition constraints, and is expressed as: ;in, Indicates time Predictive patterns, This indicates the prediction pattern from the previous moment. Indicates the current feasibility level of the forecast. This represents the current set of adaptability scores for each mode. This represents mode switching hysteresis constraints and boundary buffer constraints. This indicates the duration of the current state. This represents the state transition function.

[0066] When the model fit of each prediction mode is lower than the set threshold, or when the higher-level modes do not meet the entry conditions, the system automatically enters the conservative output state to ensure that it still has basic prediction output capability under extremely limited conditions.

[0067] S5: Construct a description of the predictive expert's capability boundary and an expert admission constraint matrix.

[0068] The prediction function unit is referred to as a prediction expert. For multiple prediction experts, the corresponding prediction expert capability boundary description information is predefined. The capability boundary description information includes at least: key observation support requirements, historical support requirements, auxiliary support requirements, operating condition disturbance tolerance range, hierarchical collaboration support requirements, task constraint adaptation range, prediction feasibility level, prediction mode, prediction level, and prediction baseline applicability range.

[0069] No. A vector describing the capabilities of a prediction expert Represented as:

[0070] ;

[0071] in, , , and They represent the first The minimum requirements for key observational support, historical support, auxiliary support, and hierarchical collaborative support for forecasting experts. and These represent the upper limits of operating condition disturbances and task constraints that the forecasting expert can tolerate, respectively. This indicates the forecasting feasibility level, forecasting model, forecasting level, and forecasting baseline range applicable to the forecasting expert.

[0072] The description information of the predictive expert's capability boundary can be determined based on preset rules in the expert design phase, statistical results of historical validation samples, cross-scenario evaluation results, or a combination thereof, thereby giving the applicable boundaries of different predictive experts a clear source basis and interpretability. Furthermore, the capability boundary can be determined at least based on the predictive expert's tolerance for missing input elements, minimum historical sample requirements, degree of dependence on auxiliary references, statistical results of boundary condition errors, and cross-scenario stability evaluation results.

[0073] Furthermore, based on the forecasting support indicators, the current forecasting feasibility level, the current target forecasting model, and the task requirements, an expert admission constraint matrix is ​​constructed to screen forecasting experts. This expert admission constraint matrix is ​​used to limit which forecasting experts are qualified and which are not under the current forecasting conditions.

[0074] Unlike conventional gating weighting methods, this step first uses a hard constraint screening of prediction experts based on the relationship between the current prediction task's supporting foundation and the prediction expert's capability boundary, rather than directly assigning weights to all prediction experts simultaneously. The admission criteria include at least the following aspects: whether the completeness of current key observations meets the minimum requirements, whether the historical sample length meets the minimum requirements, whether the auxiliary support information meets the call conditions, whether the current operational disturbance is within the prediction expert's tolerable range, and whether the current task level and prediction horizon are within their applicable range. Prediction experts that do not meet the minimum input conditions or minimum historical conditions are not included in the current prediction output calculation.

[0075] No. A forecasting expert at a certain moment The admission results are expressed as follows:

[0076] ;

[0077] in, Indicates the first A forecasting expert at a certain moment Possess the necessary qualifications. This indicates that the applicant does not meet the eligibility requirements. This indicates the prediction feasibility level of the current prediction task. This indicates the prediction mode for the current prediction task. This indicates the prediction level of the current prediction task, such as the power plant level, inverter level, string level, or other equipment level. This indicates the forecast timescale or forecast horizon of the current forecast task, such as the power plant level, inverter level, string level, or other equipment level.

[0078] Therefore, the expert admission constraint matrix at the current moment is expressed as: Based on this, a set of candidate prediction experts is formed: .

[0079] It should be noted that the expert admission process in this step does not involve directly weighting all forecasting experts. Instead, a hard screening is first implemented based on the relationship between the forecasting support indicators and the expert capability descriptions. Only forecasting experts who meet the admission criteria are allowed to enter the candidate set, thereby improving the relevance and stability of subsequent expert organization and forecast output.

[0080] S6: Organize qualified forecasting experts.

[0081] Based on the set of candidate forecasting experts formed in S5, forecasting experts with admission qualifications are organized under the current forecasting conditions, and leading forecasting experts, auxiliary forecasting experts, and conservative forecasting experts are determined.

[0082] Among them, the leading prediction expert is used to output the main prediction results under the current prediction task conditions. The auxiliary prediction expert is used to provide compensation predictions, error correction predictions, boundary scenario verification predictions, or hierarchical collaborative predictions. The conservative prediction expert is used to provide conservative prediction outputs under conditions of current prediction model degradation, instability of the leading prediction expert, continuous transition failure, or high uncertainty.

[0083] Furthermore, within the candidate prediction expert set, an expert fit score is calculated for each prediction expert under the current task. The expert fit score comprehensively considers the degree of matching between the current prediction task support foundation and the prediction expert's capability boundary, the prediction performance of the prediction expert in nearby historical scenarios, the output stability under the current target prediction mode, and the switching transition cost introduced by switching from the current dominant prediction path to the prediction expert.

[0084] No. A forecasting expert at a certain moment The expert fit score is expressed as follows: ;in, Indicates the first A forecasting expert at a certain moment Expert matching score, This indicates the current forecasting task's supporting foundation and the first The degree of matching between the descriptions of the predictive experts' abilities. Indicates the first The evaluation of a predictive expert's local performance in a nearby historical context. This represents the output stability evaluation value of the forecasting expert under the current forecasting model. This indicates a switch to the continuity acceptance cost introduced by the forecasting expert. , , and This is the adjustment coefficient.

[0085] Based on the expert fit score, the prediction expert with the highest score is selected as the lead prediction expert, denoted as:

[0086] .

[0087] From the remaining candidate prediction experts, one or more prediction experts with higher scores and complementary functions are selected as auxiliary prediction experts. "Complementary functions" mainly refer to differences between the auxiliary and leading prediction experts in terms of input dependence, historical dependence, auxiliary support dependence, boundary scenario adaptability, or hierarchical collaborative support capabilities, thereby enhancing the overall applicability under complex prediction conditions. The set of auxiliary prediction experts can be an empty set, a single-element set, or a multi-element set; when the current prediction task conditions are relatively simple or the leading prediction expert can already meet the output requirements, the system allows the leading prediction expert to output the prediction results directly without configuring auxiliary prediction experts.

[0088] Furthermore, from the set of candidate prediction experts, prediction experts with high output stability and strong continuous transition ability are selected as conservative prediction experts. Conservative prediction experts are represented as follows:

[0089] ;

[0090] in, Indicates the first The forecast continuity capability of each forecasting expert. It should be noted that under normal conditions, conservative forecasting experts do not directly participate in the generation of the main forecast result, but remain in a callable state; when the instability of the dominant forecasting expert, the failure of candidate expert switching, the degradation of the forecasting mode, or the current task enters a high uncertainty scenario is detected, the conservative forecasting expert participates in the forecast output as a backup forecasting path to ensure the continuity and basic availability of the forecast results;

[0091] In another implementation, different prediction experts may be composed of prediction models or combinations of models optimized for complete observation conditions, limited observation conditions, short historical conditions, auxiliary support conditions, or conservative output conditions.

[0092] S7: Generate initial prediction results.

[0093] The leading forecasting expert outputs the main forecasting result based on the current forecasting task, while the auxiliary forecasting expert outputs the compensation forecasting result, the verification forecasting result, the deviation correction amount, the confidence boundary or the hierarchical correction information; the initial forecasting result for the current moment is formed by combining the outputs of the leading forecasting expert and the auxiliary forecasting expert.

[0094] When the auxiliary prediction expert is used to compensate the output, its output is mainly used to correct the prediction deviation of the dominant prediction expert in scenarios with limited observation, short history, boundary conditions, or insufficient hierarchical information. When the auxiliary prediction expert is used to verify the output, its output is mainly used to compare the rationality of the dominant prediction results. When the auxiliary prediction expert is used for hierarchical correction, its output is mainly used to support the consistency constraints between the station level, inverter level, and string level.

[0095] The initial prediction result at the current moment is expressed as:

[0096] ;

[0097] in, This indicates the main forecast result output by the leading forecasting expert. This represents a set of experts assisting in prediction. Indicates the first The compensation or correction amount provided by an auxiliary forecasting expert. This represents the synergistic effect coefficient of the corresponding auxiliary prediction expert under the current conditions.

[0098] Furthermore, when auxiliary forecasting experts participate through verification rather than direct compensation, their output can be transformed into deviation confidence adjustment, anomaly boundary correction, or hierarchical allocation correction for the main forecast results, thereby enhancing the interpretability and usability of the initial forecast results under complex conditions.

[0099] It should be noted that the auxiliary prediction expert in this step does not necessarily output a complete prediction sequence. It may also output bias correction amount, confidence boundary, hierarchical quota, verification results or other collaborative information used to correct the dominant prediction results, thereby avoiding the present invention from degenerating into a common multi-model averaging or weighted ensemble scheme.

[0100] S8: Predictive Continuity Control.

[0101] Prediction continuity control is implemented on the initial prediction results and their generation path to reduce the risks of jumps and short-term oscillations in the prediction output caused by changes in prediction feasibility level, changes in target prediction mode, changes in leading prediction experts, recovery or failure of key observations, rapid changes in operating conditions, and boundary scene disturbances. The continuous transition control is used not only to reduce output abrupt changes caused by mode switching and expert switching, but also to limit the instantaneous offset of the prediction output when key observations are recovered or fail.

[0102] When implementing prediction continuity control, at least the following processes are included: (1) Continuity risk identification. Determine whether there is prediction continuity risk at the current moment. The continuity risk includes at least the following: a change in prediction feasibility level across levels, a change in the target prediction mode, a change in the leading prediction expert, a sudden change in the integrity of key observations, a significant increase in operating condition disturbances, or the initial prediction result exceeding the preset change boundary relative to the previous moment's output. (2) Parallel verification. For a new prediction expert or a new prediction path to replace the current leading prediction path, generate a reference output in parallel within a short window, and continuously verify its output stability, engineering constraint satisfaction, hierarchical consistency, and differences from the current leading prediction result. (3) Continuous transition. If the parallel verification meets the preset conditions, combine the original prediction output and the new prediction output according to the continuity control coefficient within the transition window to generate the prediction result after continuous transition. (4) Abnormal rollback. If the new prediction path experiences output instability, engineering constraint non-satisfaction, significant decrease in credibility, or deterioration in hierarchical consistency during the parallel verification stage or the continuous transition stage, the transition is stopped, and the result is rolled back to the output result of the original leading prediction expert or conservative prediction expert.

[0103] The prediction results after continuous transition are expressed as follows: ;in, This indicates the output result of the original dominant prediction path. This indicates the output of the new predicted path. This represents the continuous control coefficient, which is adjusted according to a monotonic variation pattern within the transition window. Furthermore, the continuous control coefficient... The forecast can be dynamically adjusted based on the current feasibility level, parallel verification performance, forecast model stability, degree of operational disturbance, and historical transition success rate to balance transition smoothness and response timeliness. If there is no continuity risk at the current moment, the initial forecast result can be directly used as the forecast result after continuity control.

[0104] Furthermore, to constrain the magnitude of change in the predicted output during the transition phase, the following must be satisfied: ;in, Indicates time The prediction results This indicates the prediction result at the previous moment. Indicates time The permissible change boundary is related to the current working condition scenario, the sunshine stage, the prediction level, the target object capacity, and the operating status constraints.

[0105] S9: Engineering constraint verification and result output.

[0106] The prediction results after prediction continuity control are checked against engineering constraints, and consistency corrections are made in conjunction with the hierarchical collaboration information of the current prediction task to form the final prediction results. At the same time, the credibility is evaluated and graded output based on the prediction feasibility level, expert collaboration consistency, model stability and the degree of satisfaction of engineering constraints.

[0107] The engineering constraints include at least non-negativity constraints, capacity limit constraints, near-zero output constraints at night, smooth transition constraints between sunrise and sunset, operational status constraints such as power limit, maintenance, and fault, and hierarchical aggregation consistency constraints.

[0108] When upper-level aggregated information, lower-level component information, same-level reference information, or historical structure allocation information exist, hierarchical consistency correction can be performed on the prediction results. The final prediction result is obtained from the following optimization form:

[0109] ;

[0110] in, This indicates the final prediction result. This represents the set of feasible solutions that satisfy the engineering constraints. This represents the predicted outcome variable to be optimized. This indicates the forecast result after continuous forecast control. Represents a hierarchical aggregation mapping matrix. This represents the measured value, reliable value, or reliable predicted value of the upper-layer device or object. This indicates the reference value for the same floor, the structural reference value, or the historical structural allocation reference value. and This is the adjustment coefficient.

[0111] Furthermore, the credibility assessment can comprehensively consider the following factors: the prediction feasibility level, the degree of consistency between the leading prediction expert and the auxiliary prediction experts, the stability of prediction continuity control, the degree of engineering constraint satisfaction, and the degree of improvement before and after hierarchical consistency correction. Based on the assessment results, the final prediction results can be divided into high credibility, usable for inspection, and low credibility for review levels, and the predicted value, prediction feasibility level, prediction mode, leading prediction expert identifier, auxiliary prediction expert identifier, conservative prediction expert identifier, continuity control status identifier, and credibility label can be output simultaneously.

[0112] The final prediction results can be further output to dispatching application, operation optimization, power analysis, power market trading, alarm assessment and other business application modules.

[0113] This invention proposes a photovoltaic power prediction method based on prediction feasibility classification, expert access, and prediction continuity control. It addresses power prediction tasks at the photovoltaic power plant level, inverter level, or string level. Under different conditions such as complete key observations, limited observations, insufficient historical data of the target object, available auxiliary support, and high boundary uncertainty, it uniformly quantifies the supporting foundation of the current prediction task. Based on this, prediction feasibility classification is first completed, and then the current prediction mode is determined by combining prediction mode state constraints. Subsequently, expert access screening is implemented according to the capability boundaries of prediction experts, and leading prediction experts, auxiliary prediction experts, and conservative prediction experts are organized within the access scope. Then, initial prediction results are generated, and prediction continuity control is implemented on the prediction output. Finally, the final prediction result is output by combining engineering constraint rules, hierarchical consistency verification, and credibility assessment.

[0114] The technical focus of this invention is not to limit a specific prediction model structure, but to construct an integrated prediction method that includes prediction condition quantification, prediction mode determination, prediction expert access, expert collaborative output, and continuous transition control, in order to improve the stability, continuity, and engineering usability of photovoltaic power prediction under complex conditions, taking into account changes in the integrity of key observations, the sufficiency of historical support, the availability of auxiliary support, and the degree of operating condition disturbance.

[0115] Existing photovoltaic (PV) power prediction methods lack adaptability to complex prediction conditions. These methods are typically built upon fixed input structures, fixed prediction models, or fixed prediction processes. When key observations are missing, historical samples of the target object are insufficient, operating conditions fluctuate rapidly, or multiple adverse conditions overlap, they struggle to adaptively adjust the prediction path according to the actual conditions of the current prediction task. Especially in scenarios with limited observations, short historical data, auxiliary support, or high boundary uncertainties, existing methods are prone to problems such as mismatch between the prediction model and the current task conditions, and decreased stability of prediction results, making it difficult to balance prediction accuracy, continuity, and engineering usability.

[0116] In this invention, by constructing a prediction condition feature vector and prediction support indicators, the completeness of key observations, the sufficiency of target history, the availability of auxiliary information, the volatility of operating conditions, the hierarchical collaborative support capability, and the task constraints are uniformly characterized. On this basis, the prediction feasibility classification and prediction mode determination are completed, enabling the system to adaptively select a more suitable prediction path for different prediction conditions, avoiding the mismatch problem of fixed models or fixed processes in complex scenarios.

[0117] Existing methods lack a mechanism for organizing prediction experts based on prediction feasibility grading and expert admission criteria. When key observations are missing, historical windows are insufficient, auxiliary reference information is unstable, or multiple adverse conditions overlap, fixed models or fixed prediction processes struggle to balance accuracy and stability, easily leading to mismatches between prediction paths and current task conditions. Current technologies generally lack mechanisms for limiting the admission of prediction experts based on the feasibility of the current prediction task, and for identifying leading, auxiliary, and conservative prediction experts within the admission scope.

[0118] After completing the prediction feasibility classification and prediction mode determination, this invention limits the admission of candidate prediction experts based on the capability boundary of prediction experts, and determines the leading prediction expert, auxiliary prediction expert and conservative prediction expert within the admission range. This enables different prediction experts to undertake the main prediction, compensation correction, boundary verification or conservative output tasks under more suitable prediction conditions, thereby improving the stability and applicability of photovoltaic power prediction in observation-limited, short-history, auxiliary support and boundary scenarios.

[0119] Existing methods lack control capabilities for the continuity of prediction output. During the prediction process, as key observations are recovered, historical samples are accumulated, operating conditions change abruptly, or boundary scenarios emerge, the appropriate prediction model and the leading prediction expert may change. Without a reasonable prediction continuity control mechanism, prediction jumps, short-term jitters, or unreasonable responses can easily occur at model boundaries or expert replacement boundaries, affecting the practical usability of prediction results in scheduling, control, and analysis scenarios.

[0120] In this invention, by implementing prediction continuity control, smooth constraints are applied to changes in prediction feasibility level, prediction mode, leading prediction expert, and prediction output transition process. Furthermore, by combining non-negative constraints, capacity limit constraints, nighttime near-zero constraints, sunrise and sunset transition constraints, and credibility assessment output, the risk of jumps and jitters in prediction results at mode boundaries and expert replacement boundaries is reduced, thereby improving the engineering availability of prediction results in scheduling applications, operational analysis, and business applications.

[0121] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A photovoltaic power prediction method based on hierarchical, admission, and prediction continuity control, characterized in that: Includes the following steps: S1: Collect the prediction input information corresponding to the target power station or target equipment to form the original input set for the current prediction task; S2: Based on the original input set of the current prediction task, extract the feature quantities used to characterize the prediction conditions of the current prediction task, construct the prediction condition feature vector, and calculate the prediction support index based on the prediction condition feature vector. S3: Based on the prediction condition feature vector and prediction support index, the feasibility of the current prediction task at the current moment is evaluated in a graded manner to obtain the prediction feasibility graded result. S4: Based on the prediction condition feature vector, prediction support index and prediction feasibility classification results, construct the prediction mode state machine, dynamically determine the prediction organization method suitable for the current prediction task under different conditions and working conditions, and output the prediction mode at the current moment. S5: The prediction function unit is called the prediction expert, and the description information of the prediction expert's capability boundary is defined; based on the prediction support indicators, the current prediction feasibility level, the current target prediction mode and task requirements, an expert admission constraint matrix is ​​constructed to screen prediction experts and form a candidate prediction expert set accordingly. S6: Within the set of candidate prediction experts, calculate the expert fit score for each prediction expert under the current task, select the prediction expert with the highest score as the leading prediction expert, and select one or more prediction experts with high scores and complementary functions from the remaining candidate prediction experts as auxiliary prediction experts. S7: The leading forecasting expert outputs the main forecasting result based on the current forecasting task, and the auxiliary forecasting expert outputs the compensation or correction amount; the initial forecasting result for the current moment is formed by combining the outputs of the leading forecasting expert and the auxiliary forecasting expert. S8: Implement prediction continuity control on the initial prediction results and their generation paths; S9: Perform engineering constraint verification on the prediction results after prediction continuity control, and perform consistency correction in combination with the hierarchical collaboration information of the current prediction task to form the final prediction result.

2. The photovoltaic power prediction method based on hierarchical, admission, and prediction continuity control according to claim 1, characterized in that: In S2, time Predictive conditional feature vector Represented as: ;in, Indicates key observation integrity characteristics, This indicates the sufficiency of historical support. This indicates the availability characteristics of auxiliary support. This indicates the characteristics of the operating condition disturbance. This indicates hierarchical collaborative support characteristics. Indicates task constraint characteristics; time Forecast support indicators Represented as: ;in, Represents the feature mapping function; predicts supporting indicators. Expanded by dimension: ;in, Indicates key observational support indicators, Indicates historical support indicators. This indicates a secondary support indicator. This indicates the operating condition disturbance index. Indicators representing hierarchical collaborative support. This indicates the task constraint indicators.

3. The photovoltaic power prediction method based on hierarchical, admission, and prediction continuity control according to claim 2, characterized in that: In S3, time Predictive feasibility score Represented as: ;in, This represents the adjustment coefficient for the corresponding dimension. This represents the first result obtained from the predictive support indicator. Support evaluation results Penalties indicating a combination of adverse conditions; Based on the prediction feasibility score, the current prediction task is divided into different prediction feasibility levels: ;in, Indicates time The predicted feasibility level, This represents the mapping function from the predicted feasibility score to the grade label.

4. The photovoltaic power prediction method based on hierarchical, admission, and prediction continuity control according to claim 3, characterized in that: In S4, pattern fit is represented as: ;in, Indicates the first Predictive patterns at time Mode compatibility; Indicates the relationship with the first The feature contribution coefficient vector corresponding to the prediction pattern; Indicates time The predicted conditional feature vector; Indicates the first The entry constraint violation of the prediction model under the current conditions is used to characterize the degree of non-compliance of basic conditions such as key observation lower limit, historical sample lower limit, availability of auxiliary support and stability of operating conditions. This represents the corresponding penalty coefficient; The prediction pattern at the current moment is represented as: ;in, Indicates time Predictive patterns This indicates the prediction pattern from the previous moment. Indicates the current feasibility level of the forecast. This represents the current set of adaptability scores for each mode. This represents mode switching hysteresis constraints and boundary buffer constraints. This indicates the duration of the current state. This represents the state transition function.

5. The photovoltaic power prediction method based on hierarchical, admission, and prediction continuity control according to claim 4, characterized in that: In S5, the first A vector describing the capabilities of a prediction expert Represented as: ;in, , , and They represent the first The minimum requirements for key observational support, historical support, auxiliary support, and hierarchical collaborative support for forecasting experts. and These represent the upper limits of operating condition disturbances and task constraints that the forecasting expert can tolerate, respectively. This indicates the applicable forecasting feasibility level, forecasting model, forecasting level, and forecasting baseline range for the forecasting expert. No. A forecasting expert at a certain moment The admission results are expressed as follows: ； in, Indicates the first A forecasting expert at a certain moment Possess the necessary qualifications. This indicates that the applicant does not meet the eligibility requirements. This indicates the prediction feasibility level of the current prediction task. This indicates the prediction mode for the current prediction task. This indicates the prediction level of the current prediction task. This represents the prediction timescale or prediction horizon of the current prediction task; therefore, the expert admission constraint matrix at the current moment is expressed as: Based on this, a set of candidate prediction experts is formed: .

6. The photovoltaic power prediction method based on hierarchical, admission, and prediction continuity control according to claim 5, characterized in that: In S6, the expert fit score is expressed as follows: ;in, Indicates the first A forecasting expert at a certain moment Expert matching score, This indicates the current forecasting task's supporting foundation and the first The degree of matching between the descriptions of the predictive experts' abilities. Indicates the first The evaluation of a predictive expert's local performance in a nearby historical context. This represents the output stability evaluation value of the forecasting expert under the current forecasting model. This indicates a switch to the continuity acceptance cost introduced by the forecasting expert. , , and The adjustment coefficient is used; based on the expert fit score, the predictive expert with the highest score is selected as the lead predictive expert, denoted as: 。 7. The photovoltaic power prediction method based on hierarchical, admission, and prediction continuity control according to claim 6, characterized in that: In S6, conservative prediction experts are selected from the candidate prediction expert set based on their high output stability and strong continuous transition capability. Under normal conditions, conservative prediction experts do not directly participate in the generation of the main prediction result, but remain available for use. When instability of the dominant prediction expert, failure of candidate expert switching, prediction mode degradation, or the current task entering a high-uncertainty scenario is detected, conservative prediction experts participate in the prediction output as a fallback prediction path to ensure the continuity and basic availability of the prediction results. Conservative prediction experts are represented as follows: ； in, Indicates the first The ability of a forecasting expert to maintain forecast continuity.

8. The photovoltaic power prediction method based on hierarchical, admission, and prediction continuity control according to claim 7, characterized in that: In S7, the initial prediction result at the current time is represented as: ； in, This indicates the main forecast result output by the leading forecasting expert. This represents a set of experts assisting in prediction. Indicates the first The compensation or correction amount provided by an auxiliary forecasting expert. This represents the synergistic effect coefficient of the corresponding auxiliary prediction expert under the current conditions.

9. The photovoltaic power prediction method based on hierarchical, admission, and prediction continuity control according to claim 8, characterized in that: In S8, when implementing forecast continuity control, it is first determined whether there is a forecast continuity risk at the current moment. Then, parallel verification is performed on the new forecast expert or new forecast path that is to replace the current dominant forecast path. If the parallel verification meets the preset conditions, the original forecast output and the new forecast output are combined according to the continuity control coefficient within the transition window to generate the forecast result after continuous transition. If the parallel verification does not meet the preset conditions, it reverts to the output result of the original dominant forecast expert or conservative forecast expert. In S8, the prediction result after continuous transition is expressed as: ;in, This indicates the output result of the original dominant prediction path. This indicates the output of the new predicted path. This represents the continuous control coefficient, which is adjusted according to a monotonic variation pattern within the transition window; To constrain the magnitude of change in the predicted output during the transition phase, the following must be satisfied: ;in, Indicates time The prediction results This indicates the prediction result at the previous moment. Indicates time The permissible change boundary.

10. The photovoltaic power prediction method based on hierarchical, admission, and prediction continuity control according to claim 9, characterized in that: In S9, the final prediction result is obtained by the following optimization: ； in, This indicates the final prediction result. This represents the set of feasible solutions that satisfy the engineering constraints. This represents the predicted outcome variable to be optimized. This indicates the forecast result after continuous forecast control. Represents a hierarchical aggregation mapping matrix. This represents the measured value, reliable value, or reliable predicted value of the upper-layer device or object. This indicates the reference value for the same floor, the structural reference value, or the historical structural allocation reference value. and This is the adjustment coefficient.

Images