Method and device for persistent anchoring of hydroelectric day-ahead power forecasts, storage medium

Abstract

The embodiment of the application discloses a kind of water and electricity day-ahead power prediction method and device of persistent anchoring, storage medium, the method comprises the following steps: S1, the historical water and electricity power sequence of the water power station to be predicted is acquired, and historical meteorological sequence and target day meteorological sequence are time-aligned with the historical water and electricity power sequence;S2, according to the latest operation window of the historical water and electricity power sequence Calculation persistence anchor value;S3, construct and train to obtain persistent anchoring transition deviation learning model;S4, the next day average active power prediction value is obtained by the learning model and is output by prediction.The present application processes regular, moderate transition and extreme transition as different correction paths, so that different severity weather-driven deviations obtain different correction permissions.The present application can retain the reliable reference provided by recent power level in regular day;At the same time, moderate and extreme signed residual provide directional deviation correction, which is beneficial to reduce transition day prediction error.

Description

Technical Field

[0001] This invention relates to the field of hydropower power prediction and power system dispatching technology, and in particular to a persistently anchored method, apparatus, and storage medium for predicting hydropower day-ahead power. Background Technology

[0002] Reliable next-day power forecasts are needed for day-ahead power generation planning, dispatch preparation, and reservoir coordination for run-of-river hydropower stations. For most routine operating days, hydropower output typically varies smoothly with recent operating conditions, and persistent references based on recent power levels can provide relatively stable forecasts.

[0003] Under conditions of rainfall, snowmelt, or rapid weather changes, upstream water inflow and generator output may deviate significantly the following day. While the number of samples of this type of weather-driven shift is relatively small, it often contributes a significant amount to scheduling-related forecast errors. Furthermore, the intensity of deviation varies among different shift samples, ranging from moderate directional corrections to stronger extreme deviations.

[0004] Existing general-purpose time series prediction networks can enhance representation learning capabilities, but if a single regression head is used to directly output predicted values, it is easy to mix regular samples with transition samples of different severity. It is difficult to assign different strengths of correction authority to moderate and extreme transitions while preserving the reliability of regular daily persistence, resulting in increased error. Summary of the Invention

[0005] The technical problem to be solved by the embodiments of the present invention is to provide a persistently anchored method, apparatus and storage medium for predicting hydropower day-ahead power, so as to reduce the prediction error of the transition day.

[0006] To address the aforementioned technical problems, this invention proposes a persistently anchored method for predicting day-ahead hydropower, comprising the following steps: S1, obtain the historical hydropower power sequence of the hydropower station to be predicted, the historical meteorological sequence time-aligned with the historical hydropower power sequence, and the meteorological sequence of the target day; S2, calculate the persistence anchor value based on the most recent operating window of the historical hydropower power sequence; S3, construct and train a persistent anchoring shift bias learning model; S4, the predicted value of the average active power for the next day is obtained through the learning model and output; The learning model includes a shared hydropower-meteorological context encoder, a hierarchical transformation router, a signed residual generator, and a hierarchical sensing predictive synthesizer. The context encoder is used to generate a shared hydropower-meteorological context, a hydropower status summary, and a hydrometeorological driving summary. The hierarchical transformation router obtains the normal level probability, the medium level probability, the extreme level probability, and the activation strength of the medium and extreme transformation paths based on the shared hydropower-meteorological context, the hydropower state summary, the hydrometeorological driving summary, and the persistent anchor value. The signed residual generator generates a regular correction based on the shared hydro-meteorological context and the hydro-state summary, and generates intermediate-transition signed residuals and extreme-transition signed residuals based on the output of the hierarchical transformation router, respectively. The hierarchical sensing prediction synthesizer performs hierarchical sensing prediction synthesis based on the persistent anchor value, the conventional correction amount, the intermediate transition signed residual, the extreme transition signed residual, and the corresponding activation intensity to obtain the predicted value of the average active power for the next day.

[0007] Accordingly, embodiments of the present invention also provide a persistently anchored hydropower day-ahead power prediction device, comprising: Data acquisition module: used to acquire the historical hydropower output sequence, historical meteorological sequence, and target day meteorological sequence of the hydropower station to be predicted; Anchor value calculation module: used to calculate persistent anchor values ​​based on the most recent running window, and to determine the normal level, medium transformation level, and extreme transformation level; A shared coding module is used to generate shared hydropower-meteorological context, hydropower status summary, and hydrometeorological driving summary; The hierarchical routing module is used to generate regular hierarchical probabilities, medium transition hierarchical probabilities, extreme transition hierarchical probabilities, and the activation strength of medium transition paths and extreme transition paths. The residual generation module is used to generate conventional correction amounts, medium-transformation signed residuals, and extreme-transformation signed residuals. The prediction synthesis module obtains and outputs the predicted value of the average active power for the next day through hierarchical perception and prediction synthesis.

[0008] Accordingly, embodiments of the present invention also provide a computer storage medium storing a computer program, which, when executed by a processor, implements the method described above.

[0009] The beneficial effects of this invention are as follows: compared with the method of inputting all samples into a single regression head, this invention treats normal, moderate and extreme changes as different correction paths, so that weather-driven deviations of different severity can obtain different correction permissions.

[0010] Since the final prediction is synthesized around a persistent anchor value, this invention is able to retain a reliable reference provided by the recent power levels in the normal day; at the same time, the moderate and extreme signed residuals provide directional deviation corrections, which helps to reduce the prediction error of the transition day. Attached Figure Description

[0011] Figure 1 This is a schematic flowchart of the persistently anchored hydropower day-ahead power prediction method according to an embodiment of the present invention.

[0012] Figure 2 This is a schematic diagram of the structure of the persistent anchoring shift bias learning model according to an embodiment of the present invention.

[0013] Figure 3 This is a flowchart illustrating the persistently anchored hydropower day-ahead power prediction method of Embodiment 1 of the present invention. Detailed Implementation

[0014] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] In this embodiment of the invention, directional indicators (such as up, down, left, right, front, back, etc.) are only used to explain the relative positional relationship and movement of each component in a specific posture (as shown in the figure). If the specific posture changes, the directional indicator will also change accordingly.

[0016] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features.

[0017] Please refer to Figure 1 The persistently anchored hydropower day-ahead power prediction method of this invention includes the following steps S1 to S4.

[0018] Step S1: Obtain the historical hydropower power sequence of the hydropower station to be predicted, the historical meteorological sequence time-aligned with the historical hydropower power sequence, and the meteorological sequence for the target day.

[0019] Step S2: Calculate the persistent anchor value based on the most recent operating window of the historical hydropower power sequence. Then, determine the normal level, moderate transition level, and extreme transition level based on the difference between the actual power in the training samples and the persistent anchor value.

[0020] Step S3: Construct and train a persistent anchoring shift bias learning model.

[0021] Step S4: The predicted value of the average active power for the next day is obtained through the learning model and output.

[0022] like Figure 2As shown, the learning model includes a shared hydropower-meteorological context encoder, a hierarchical transformation router, a signed residual generator, and a hierarchical sensing predictive synthesizer.

[0023] The context encoder is used to generate a shared hydropower-meteorological context, a hydropower status summary, and a hydrometeorological driving summary. The context encoder converts the historical hydropower power sequence into a hydropower patch representation, converts the historical meteorological sequence and the target day's meteorological sequence into meteorological token representations, and performs shared hydropower-meteorological context encoding on the hydropower patch representation and the meteorological token representation to obtain the shared hydropower-meteorological context, the hydropower status summary, and the hydrometeorological driving summary.

[0024] The hierarchical transformation router obtains the normal level probability, the medium level probability, the extreme level probability, and the activation strength of the medium and extreme transformation paths based on the shared hydropower-meteorological context, the hydropower state summary, the hydrometeorological driving summary, and the persistent anchor value.

[0025] The signed residual generator generates a regular correction based on the shared hydro-meteorological context and the hydro-state summary, and generates intermediate-transformation signed residuals and extreme-transformation signed residuals based on the output of the hierarchical transformation router.

[0026] The hierarchical sensing predictive synthesizer performs hierarchical sensing predictive synthesis based on the persistent anchor value, the conventional correction amount, the intermediate transition signed residual, the extreme transition signed residual, and the corresponding activation intensity to obtain the predicted value of the average active power for the next day.

[0027] In one implementation, in step S2, the persistence anchor value is the average value of the most recent M moments of the historical hydropower power sequence.

[0028] ; Where p^(i) represents the persistent anchor value of the i-th training sample; M represents the length of the most recent running window, i.e. the number of recent historical moments used to calculate the average; m represents the reverse time index within the window; L represents the length of the historical hydropower power sequence or the historical end position corresponding to the prediction start point; This represents the observed hydropower active power of the i-th sample at the m-th historical moment before the prediction start point. The formula uses the average of the most recent M historical power values ​​to form a regular reference level for the next day's prediction.

[0029] During the training phase, the normal level, medium transition level, and extreme transition level are determined based on the difference between the actual power in the training samples and the persistent anchor value. The difference is the absolute difference between the actual daily average active power of the training samples and the persistent anchor value. When the difference is less than a preset medium transition threshold, it is determined to be at the normal level; when the difference is greater than or equal to the preset medium transition threshold and less than a preset extreme transition threshold, it is determined to be at the medium transition level; and when the difference is greater than or equal to the preset extreme transition threshold, it is determined to be at the extreme transition level. The formula is as follows: ; Where g^(i) represents the anchor value deviation of the i-th sample; y^(i) represents the actual average active power of the i-th sample on the next day; p^(i) represents the corresponding persistent anchor value; and |·| represents the absolute value operation. Based on the relationship between g^(i) and the preset medium transition threshold and the preset extreme transition threshold, the samples are divided into normal level, medium transition level, or extreme transition level.

[0030] In one implementation, the learning model performs patch embedding on the historical hydropower power sequence to obtain a hydropower patch token; performs token embedding on the concatenated sequence of the historical meteorological sequence and the meteorological sequence of the target day to obtain a meteorological token; concatenates the hydropower patch token and the meteorological token and inputs them into the context encoder (shared Transformer encoder), and pools the encoding result of the context encoder to obtain the shared hydropower-meteorological context.

[0031] ; Wherein, H^(i) represents the shared hidden state sequence output by the shared hydropower-meteorological context encoder for the i-th sample; The parameter is The shared Transformer encoder; Z_x^(i) represents the hydropower patch token obtained from the historical hydropower power sequence; Z_w^(i) represents the meteorological token obtained from the historical meteorological sequence and the meteorological sequence of the target day; [·;·] indicates concatenation by token dimension. This formula is used to form the shared context required for subsequent routing and residual generation.

[0032] The hydropower status summary is obtained from the hydropower-side coded status, the mean of the most recent operating window, and the mean difference between the most recent operating window and the previous operating window. The hydrometeorological driving summary is obtained from the meteorological-side coded status and the meteorological statistics of the target day. The formula is as follows: ; in, This represents a summary of the hydroelectric status of the i-th sample; φ s (·) represents the hydropower state summary mapping function; Pool x (·) indicates pooling the hydropower side tokens in the shared hidden state; H^(i) represents the sequence of shared hidden states; (x^(i)) represents the mean of the most recent M historical hydropower values ​​of the i-th sample; (x^(i)) represents the mean difference between the most recent running window and the previous running window; [·;·;·] represents feature concatenation. Hydrometeorological driven summaries can be similarly formed from meteorological latent state pooling results and meteorological statistics for the target day.

[0033] In one implementation, the hierarchical transformation router generates a hierarchical logit based on routing features consisting of the shared hydropower-meteorological context, the hydropower state summary, the hydrometeorological driving summary, and the persistent anchor value, and obtains the normal hierarchical probability, the medium transformation hierarchical probability, and the extreme transformation hierarchical probability through softmax operation.

[0034] As one implementation, the hierarchical transition router further generates moderately open logits and extremely open logits based on the routing characteristics, and performs budget normalization on the pre-opening strength of the moderate and extremely open transition paths by combining the learnable boundary, regular lock score, and transition budget to obtain the opening strength. The corresponding formula is as follows: ; in, , and represents the probability that the i-th sample belongs to the normal level, the intermediate transition level, and the extreme transition level, respectively; softmax(·) represents the normalized exponential function; q^(i) represents the routing feature composed of shared hydropower-meteorological context, hydropower status summary, hydrometeorological driving summary, and persistent anchor value; W π and b π represents the weight matrix and bias term of the hierarchical logit, respectively; r, m, and e correspond to the three hierarchical types of normal, moderate, and extreme transformation, respectively.

[0035] ; in, α represents the final opening strength of the transition path in the i-th sample; β^(i) represents the allocatable transition budget obtained from the regular lock score and transition budget; α ~m ^(i) and α ~e ^(i) represent the pre-opening strength for the moderate and extreme transition paths, respectively; ε represents a small positive constant to prevent the denominator from being zero. Extreme transition path opening strength α e^(i) can be obtained using the same budget normalization method, thereby suppressing unnecessary strong corrections in regular samples.

[0036] As one implementation, generating the intermediate-transition signed residual and the extreme-transition signed residual includes: generating the magnitude and direction logit for the corresponding level, and multiplying the magnitude by the hyperbolic tangent of the direction logit to obtain the signed residual for the corresponding level. The corresponding formula is as follows: ; in, Let represent the signed residual of the intermediate transformation of the i-th sample; This indicates the magnitude of the medium-level transformation correction, used to control the amount of correction. The logit represents the direction of the intermediate transformation, used to control the correction direction; tanh(·) represents the hyperbolic tangent function, restricting the direction term to the range [-1, 1]. The extreme transformation has a signed residual Δ. e ^(i) can be modified by extreme transformation of amplitude a e ^(i) and extreme change direction logit d e ^(i) is obtained in the same way.

[0037] In one implementation, the hierarchical sensing predictive synthesizer adds the persistent anchor value to the conventional correction amount weighted by the conventional security budget, and further superimposes the moderate transition signed residual weighted by the moderate transition path activation strength and the extreme transition signed residual weighted by the extreme transition path activation strength. The corresponding formula is as follows: ; in, p^(i) represents the predicted average active power for the next day for the i-th sample; p^(i) represents the persistent anchor value. This represents the standard security budget or standard path weight; Indicates the standard correction amount; and These represent the activation strength of the intermediate and extreme transition paths, respectively. and represents the signed residuals of moderate transitions and extreme transitions, respectively. This formula combines the persistent reference, regular minor corrections, and directional transition corrections of varying severity into the final prediction.

[0038] As one implementation method, in step S3, during training, the anchor residual is decomposed into a regular residual target, a medium-transformation residual target, and an extreme-transformation residual target. Based on the weighted sum of prediction loss, hierarchical classification loss, activation loss, residual regression loss, orientation loss, and hard transformation refinement loss, the shared hydropower-meteorological context encoder, hierarchical transformation router, signed residual generator, and hierarchical sensing prediction synthesizer are jointly updated.

[0039] ; Where r^(i) represents the anchor residual of the i-th sample; y^(i) represents the actual average active power of the next day; and p^(i) represents the persistent anchor value. During training, the anchor residual is further truncated and decomposed into regular residual targets, moderately transformed residual targets, and extremely transformed residual targets, so that different deviation intensities correspond to different learning paths.

[0040] ; Where L represents the total loss of joint training; L pred Indicates the primary forecast loss; L tier L represents hierarchical classification loss; open L represents the path opening loss; res L represents the residual regression loss; dir Indicates directional loss; L hard Indicates the hard transition refinement loss; λ tier , λ open , λ res , λ dir and λ hard These represent the weight coefficients of the corresponding loss terms. This joint objective is used to simultaneously update the shared encoder, hierarchical transformation router, signed residual generator, and predictive synthesizer.

[0041] Please refer to Figure 2 The persistently anchored hydropower day-ahead power prediction device according to an embodiment of the present invention includes: Data acquisition module: used to acquire the historical hydropower output sequence, historical meteorological sequence, and target day meteorological sequence of the hydropower station to be predicted; Anchor value calculation module: used to calculate persistent anchor values ​​based on the most recent running window, and to determine the normal level, medium transformation level, and extreme transformation level; A shared coding module is used to generate shared hydropower-meteorological context, hydropower status summary, and hydrometeorological driving summary; The hierarchical routing module is used to generate regular hierarchical probabilities, medium transition hierarchical probabilities, extreme transition hierarchical probabilities, and the activation strength of medium transition paths and extreme transition paths. The residual generation module is used to generate conventional correction amounts, medium-transformation signed residuals, and extreme-transformation signed residuals. The prediction synthesis module obtains and outputs the predicted value of the average active power for the next day through hierarchical perception and prediction synthesis.

[0042] Each module of the present invention can be implemented by one or more processors executing program instructions in memory, or it can be deployed in a hydropower station dispatching system, edge computing device or cloud server.

[0043] The data acquisition module is used to acquire historical hydropower power sequences, historical meteorological sequences, and target day meteorological sequences; the anchor value calculation module is used to calculate persistent anchor values; the shared coding module is used to output shared hydropower-meteorological context, hydropower status summary, and hydrometeorological driving summary; the hierarchical routing module is used to output hierarchical probabilities and activation intensity; the residual generation module is used to output conventional corrections, moderate transition signed residuals, and extreme transition signed residuals; and the prediction synthesis module is used to output the next day's average active power prediction value.

[0044] The computer storage medium of this invention stores a computer program, which, when executed by a processor, implements a persistently anchored hydropower day-ahead power prediction method.

[0045] Example 1: As Figure 3 As shown in this embodiment, a persistently anchored method for predicting day-ahead hydropower is provided. This method is applicable to run-of-river hydropower stations or other hydropower stations with significant inflow and weather-driven effects, and the predicted output is the average active power forecast for the next day.

[0046] Step S1: Obtain the historical hydropower output sequence of the hydropower station to be predicted, the historical meteorological sequence time-aligned with the historical hydropower output sequence, and the meteorological sequence for the target day. The historical hydropower output sequence is used to describe the recent operating level, trend, and short-term fluctuations of the power station, while the historical meteorological sequence and the meteorological sequence for the target day are used to describe the hydrometeorological drivers affecting future inflow response.

[0047] Step S2: Calculate the persistent anchor value based on the most recent operating window of the historical hydropower power sequence. This persistent anchor value represents the normal reference power level under the current operating conditions. During the training phase, the samples can be divided into normal level, moderate transition level, and extreme transition level based on the degree of deviation between the actual next day's average active power and the persistent anchor value.

[0048] Step S3 involves encoding a shared hydropower-meteorological context. Specifically, historical hydropower power sequences are segmented and embedded into hydropower patch tokens, and historical meteorological sequences and target day meteorological sequences are concatenated and embedded into meteorological tokens. The two types of tokens are concatenated and input into a shared Transformer encoder to obtain shared hidden states, which are then pooled to obtain the shared hydropower-meteorological context. Further, a hydropower state summary is obtained based on the hydropower-side hidden states, the mean of the most recent operating window, and the difference between the window means, while a hydrometeorological-driven summary is obtained based on the meteorological-side hidden states and the meteorological statistics of the target day.

[0049] Step S4: The shared hydropower-meteorological context, hydropower state summary, hydrometeorological driving summary, and persistent anchor values ​​are combined to form routing features, which are then input into the hierarchical transformation router. The hierarchical transformation router outputs regular hierarchical probabilities, moderate transformation hierarchical probabilities, and extreme transformation hierarchical probabilities, and generates the activation strength of moderate and extreme transformation paths. Preferably, the activation strength is constrained by the regular lock score and transformation budget to suppress unnecessary strong corrections on regular samples.

[0050] Step S5: Generate conventional corrections based on the shared hydropower-meteorological context and hydropower state summary, and generate intermediate-transformation signed residuals and extreme-transformation signed residuals based on the output of the hierarchical transformation router. Both intermediate and extreme signed residuals can be formed by the corresponding correction magnitude and direction logit, thus simultaneously expressing the correction strength and correction direction.

[0051] Step S6 involves performing hierarchical sensing prediction synthesis based on persistent anchor values, conventional correction amounts, signed residuals of moderate transitions, signed residuals of extreme transitions, and corresponding onset intensities. This synthesis uses persistent anchor values ​​as the conventional reference path and superimposes directional corrections of varying severity based on the hierarchical routing results.

[0052] Step S7: Output the predicted average active power for the next day. This predicted value can be used for hydropower station day-ahead generation planning, dispatch preparation, output risk assessment, or electricity market participation.

[0053] Example 2, Model training method of the present invention: During the training phase, the anchor residual is first formed based on the difference between the actual average active power of the next day and the persistent anchor value. This anchor residual can be decomposed into a regular residual target, a moderate transition residual target, and an extreme transition residual target. The regular residual target corresponds to a small correction around the persistent anchor value, while the moderate and extreme residual targets correspond to stronger transition deviations.

[0054] The joint training objectives of the model include master prediction loss, hierarchical classification loss, onset loss, residual regression loss, orientation loss, and hard transformation refinement loss. The master prediction loss constrains the final predicted value, the hierarchical classification loss constrains the route probability, the onset loss constrains whether neutral and extreme paths are enabled, the residual regression loss constrains the value of each residual head, the orientation loss constrains the orientation of the signed residuals, and the hard transformation refinement loss further emphasizes extreme hierarchical samples.

[0055] Through the aforementioned joint objective, the shared hydropower-meteorological context encoder, hierarchical transition router, signed residual generator, and hierarchical sensing predictive synthesizer can be updated simultaneously, enabling transition identification, corrective permission allocation, and final power prediction to be co-optimized in the same training process.

[0056] Example 3, Verification Example: In this validation example, the data comes from a run-of-river hydropower station in Qingping Town, Mianzhu, Sichuan Province, and includes hourly active power from January 2024 to August 2025 and seven local weather variables. The seven local weather variables include air pressure, precipitation, temperature, instantaneous wind, wind speed, visibility, and relative humidity. The objective is to directly predict the average active power for the next day. The test months can be set to September 2024, November 2024, and February 2025 to cover the flood season, the normal operating period, and the low-output season.

[0057] In the principal component analysis, using Persistence as the persistence benchmark, the PTDL method corresponding to this invention reduced the overall MAE from 1.4316 to 1.4027, an absolute reduction of 0.0289, representing an improvement of approximately 2.0%; and reduced the MAE of the transition sample from 5.5275 to 4.5152, an absolute reduction of 1.0123, representing a reduction of approximately 18.3%. The transition sample refers to samples where the difference between the actual next-day average active power and the persistence anchor value reaches or exceeds the medium transition threshold.

[0058] Furthermore, in rolling validation, PTDL improved the overall MAE by 6.3% relative to Persistence, while reducing the transition MAE by 18.0%. Subtractive ablation revealed that removing shared hydro-meteorological context coding, hierarchical transition routing, signed residual separation, or hierarchical sensing-predictive synthesis control all compromised the overall error or transition error. The validation data above demonstrate that this invention, while maintaining the stability of conventional daily persistence references, can provide directional corrections for weather-driven moderate and extreme transition samples.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A persistently anchored method for predicting the day-ahead power of hydropower, characterized in that, Includes the following steps: S1, obtain the historical hydropower power sequence of the hydropower station to be predicted, the historical meteorological sequence time-aligned with the historical hydropower power sequence, and the meteorological sequence of the target day; S2, calculate the persistent anchor value based on the most recent operating window of the historical hydropower power sequence; S3, construct and train a persistent anchoring shift bias learning model; S4, the predicted value of the average active power for the next day is obtained through the learning model and output; The learning model includes a shared hydropower-meteorological context encoder, a hierarchical transformation router, a signed residual generator, and a hierarchical sensing predictive synthesizer. The context encoder is used to generate a shared hydropower-meteorological context, a hydropower status summary, and a hydrometeorological driving summary. The hierarchical transformation router obtains the normal level probability, the medium level probability, the extreme level probability, and the activation strength of the medium and extreme transformation paths based on the shared hydropower-meteorological context, the hydropower state summary, the hydrometeorological driving summary, and the persistent anchor value. The signed residual generator generates a regular correction based on the shared hydro-meteorological context and the hydro-state summary, and generates intermediate-transition signed residuals and extreme-transition signed residuals based on the output of the hierarchical transformation router, respectively. The hierarchical sensing prediction synthesizer performs hierarchical sensing prediction synthesis based on the persistent anchor value, the conventional correction amount, the intermediate transition signed residual, the extreme transition signed residual, and the corresponding activation intensity to obtain the predicted value of the average active power for the next day.

2. The persistently anchored day-ahead hydropower prediction method as described in claim 1, characterized in that, In step S2, the persistent anchor value is the average value of the historical hydropower power sequence over the most recent several moments. During the training phase, the normal level, medium transition level, and extreme transition level are determined based on the difference between the actual power in the training samples and the persistent anchor value. The difference is the absolute difference between the actual daily average active power of the training samples and the persistent anchor value. When the difference is less than a preset medium transition threshold, it is determined to be the normal level. When the difference is greater than or equal to the preset medium transition threshold and less than the preset extreme transition threshold, it is determined to be the medium transition level. When the difference is greater than or equal to the preset extreme transition threshold, it is determined to be the extreme transition level.

3. The persistently anchored day-ahead hydropower prediction method as described in claim 1, characterized in that, The learning model performs patch embedding on the historical hydropower power sequence to obtain a hydropower patch token; performs token embedding on the concatenated sequence of the historical meteorological sequence and the meteorological sequence of the target day to obtain a meteorological token; concatenates the hydropower patch token and the meteorological token and inputs them into the context encoder, and pools the encoding result of the context encoder to obtain the shared hydropower-meteorological context. The hydropower status summary is obtained from the hydropower-side coded status, the average value of the most recent operating window, and the difference between the average values ​​of the most recent operating window and the previous operating window. The hydro-meteorological driving summary is obtained from the meteorological-side coded status and the meteorological statistics of the target day.

4. The persistently anchored day-ahead hydropower prediction method as described in claim 1, characterized in that, The hierarchical transformation router generates a hierarchical logit based on the routing features consisting of the shared hydropower-meteorological context, the hydropower state summary, the hydrometeorological driving summary, and the persistent anchor value, and obtains the normal hierarchical probability, the medium transformation hierarchical probability, and the extreme transformation hierarchical probability through softmax operation.

5. The persistently anchored day-ahead hydropower prediction method as described in claim 4, characterized in that, The hierarchical transition router also generates moderately open logit and extremely open logit based on the routing characteristics, and performs budget normalization on the pre-opening strength of the moderate transition path and the extremely transition path by combining the learnable boundary, the regular lock score and the transition budget to obtain the opening strength.

6. The persistently anchored day-ahead hydropower prediction method as described in claim 1, characterized in that, Generating the intermediate-transition signed residual and the extreme-transition signed residual includes: generating the magnitude and direction logit of the corresponding level, and multiplying the magnitude by the hyperbolic tangent of the direction logit to obtain the signed residual of the corresponding level.

7. The persistently anchored day-ahead hydropower prediction method as described in claim 1, characterized in that, The hierarchical sensing predictive synthesizer adds the persistent anchor value to the conventional correction amount weighted by the conventional security budget, and further superimposes the moderate transition signed residual weighted by the moderate transition path opening strength and the extreme transition signed residual weighted by the extreme transition path opening strength.

8. The persistently anchored day-ahead hydropower prediction method as described in claim 1, characterized in that, In step S3, during training, the anchor residual is decomposed into regular residual targets, medium-transformation residual targets, and extreme-transformation residual targets. Based on the weighted sum of prediction loss, hierarchical classification loss, onset loss, residual regression loss, orientation loss, and hard transformation refinement loss, the shared hydropower-meteorological context encoder, hierarchical transformation router, signed residual generator, and hierarchical sensing prediction synthesizer are jointly updated.

9. A persistently anchored hydropower day-ahead power prediction device, characterized in that, include: Data acquisition module: used to acquire the historical hydropower output sequence, historical meteorological sequence, and target day meteorological sequence of the hydropower station to be predicted; Anchor value calculation module: used to calculate persistent anchor values ​​based on the most recent running window, and to determine the normal level, medium transformation level, and extreme transformation level; A shared coding module is used to generate shared hydropower-meteorological context, hydropower status summary, and hydrometeorological driving summary; The hierarchical routing module is used to generate regular hierarchical probabilities, medium transition hierarchical probabilities, extreme transition hierarchical probabilities, and the activation strength of medium transition paths and extreme transition paths. The residual generation module is used to generate conventional correction amounts, medium-transformation signed residuals, and extreme-transformation signed residuals. The prediction synthesis module obtains and outputs the predicted value of the average active power for the next day through hierarchical perception and prediction synthesis.

10. A computer storage medium, characterized in that, The medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images