A semantic driving-based active hierarchical self-adaptive wind power prediction method and system

Abstract

This invention relates to the field of wind power prediction technology, and discloses a semantically driven, proactive, hierarchical, adaptive wind power prediction method and system. It transforms wind power prediction from a traditional fixed, one-time regression process into a multi-stage dynamic reasoning process under input correlation, risk-driven mechanisms, and semantic constraints. First, an initial prediction and risk characterization of the future power trajectory are obtained through a global prediction module. Then, candidate high-risk intervals are formed through future-step risk scoring. Subsequently, high-level semantic context is generated using semantic reprogramming and conditional modeling. In each round of hierarchical reasoning, key prediction intervals, key driving factors, and whether to continue correction are jointly selected. Finally, local incremental correction is performed only on the selected key intervals until a stopping condition is met, at which point the final prediction result is output. This invention improves the accuracy, stability, and interpretability of wind power prediction in complex and fluctuating scenarios.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, specifically to a semantically driven active hierarchical adaptive wind power prediction method and system. Background Technology

[0002] In recent years, with the continuous growth of new energy installed capacity and the increasing demand for coordinated operation of power generation, grid, load, and storage, wind power, photovoltaic power, and net load forecasting have become crucial foundations for supporting the safe dispatching, reserve configuration, real-time control, and operation of the power market. Especially with a high proportion of renewable energy integrated into the grid, wind power is significantly affected by factors such as wind speed, wind direction, temperature, humidity, atmospheric pressure, terrain, and unit operating status, exhibiting marked randomness, volatility, and non-stationarity. Forecasting errors not only reduce the economic efficiency of dispatching plans but may also further increase frequency control pressure, reserve capacity requirements, and system operational risks.

[0003] Regarding the technical issues of improving the accuracy of wind power prediction, existing technologies can be roughly divided into the following categories.

[0004] The first category is prediction methods based on physical and statistical models. Physical models are typically built based on weather forecasts, wind turbine power curves, aerodynamic relationships, and station operating parameters. They can demonstrate a certain degree of physical interpretability, but are quite sensitive to wind field topography, weather forecast accuracy, unit parameters, and environmental disturbances. Statistical models, such as autoregressive models, Kalman filter models, and support vector regression models, make predictions through the statistical relationships between historical power and meteorological variables. They have the advantages of relatively simple structure and low computational cost, but their expressive power is limited when dealing with strongly nonlinear, multivariate coupling, and complex fluctuation events.

[0005] The second category is wind power prediction methods based on deep learning. With the development of Long Short-Term Memory (LSTM) networks, Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), Transformers, and their variants, deep learning methods have gradually become an important technical approach for wind power prediction. These methods can learn complex nonlinear mapping relationships and long- and short-term dependencies from multivariate historical sequences, and typically exhibit stronger fitting capabilities compared to traditional statistical models.

[0006] The third category comprises multi-scale decomposition, operating condition awareness, and probabilistic prediction methods. These methods model the multi-timescale fluctuations, uncertainties, and operational state differences in wind power sequences. For example, multi-scale decomposition extracts low-frequency trends and high-frequency disturbances, operating condition identification characterizes state transitions under different wind conditions, or probabilistic prediction outputs intervals, distributions, and scenarios to enhance the model's ability to characterize uncertainties. While these methods improve prediction stability to some extent, most still assume that the entire future prediction interval needs to be processed uniformly, and have not yet fully addressed the questions of "which future periods require priority correction" and "which variables dominate the prediction error during those periods."

[0007] The fourth category includes methods related to time series foundational models, cue learning, and semantic reprogramming. Existing research shows that by mapping time series segments to a higher-level semantic space and using task cues, domain knowledge, and statistical context to guide model reasoning, the model's ability to express complex evolutionary patterns can be enhanced. Relevant literature mentions methods such as Time-LLM, which reprogram time series segments to a text prototype space and use cues as context prefixes, demonstrating that time series prediction can achieve stronger generalization and reasoning capabilities in a higher-level semantic space. However, in the scenario of power source and load prediction, a complete technical solution is still lacking for unifying semantic representation, uncertainty output, key interval localization, local correction triggering, and stopping condition judgment into an implementable prediction process.

[0008] The inventors compiled currently available literature on relevant physical models and statistical predictions, including literature on wind power prediction using LSTM, Transformer, and graph neural networks; literature on multi-scale decomposition and wind condition identification; literature on probabilistic prediction and diffusion generation; and literature on basic time series models and representative prediction models such as Time-LLM, PatchTST, TimesNet, Autoformer, and iTransformer. Further systematic analysis revealed that while currently available wind power prediction and multivariate time series prediction methods have improved in terms of modeling capabilities and prediction accuracy, they still have the following shortcomings in scenarios with strong fluctuations in renewable energy.

[0009] First, existing methods mostly employ a one-time global prediction paradigm with a fixed computation graph, lacking a proactive computation allocation mechanism for high-risk future periods. Regardless of whether a period in the future sequence is characterized by steady change or drastic abrupt change, the model typically makes predictions in the same uniform manner, making it difficult to concentrate limited computational resources on critical segments such as ramp-up, sudden drops, peak-to-valley transitions, and weather shifts. Therefore, such methods are prone to significant errors in the few critical periods that truly impact scheduling safety and economy.

[0010] Second, existing adaptive methods mostly remain at the model structure level, rather than at the proactive decision-making level of the prediction process. While multi-scale fusion, dynamic weights, or condition-aware models can enhance feature representation capabilities, they typically do not explicitly determine "which segment of the future trajectory in the current sample most needs priority correction," nor do they explicitly determine "which meteorological variables, historical power inertia, or time context dominate the prediction error during this critical period." Therefore, existing methods primarily enhance overall modeling capabilities rather than achieving targeted inference for critical risk intervals.

[0011] Third, there is a lack of effective connection between existing uncertainty predictions and high-level decision-making. Probabilistic prediction methods can output prediction intervals, distributions, or confidence levels, but this information is usually only used as supplementary descriptions of the prediction results and has not been fully transformed into local correction trigger signals, key interval location criteria, and dynamic stopping conditions. Therefore, although the model can perceive uncertainty, it is difficult to further use uncertainty to guide the adaptive adjustment of the prediction process.

[0012] Fourth, existing temporal semantic enhancement methods have not fully integrated domain knowledge and key event patterns in power source-load forecasting. Some basic temporal models or cue learning methods can improve generalization ability, but they usually lack semantic condition construction for specific power scenarios such as wind power ramp-up, sudden drop, weather changes, and changes in net source-load. They also lack a unified mechanism to simultaneously use semantic conditions for interval selection, factor selection, local correction, and stopping judgment.

[0013] Fifth, the existing methods lack interpretability in their prediction process. Traditional deep learning prediction models typically output future power sequences directly, making it difficult to explain which future intervals the errors mainly originate from, which driving variables are of particular concern, and why the model continues or stops correcting. This presents certain limitations in power operations that require strong interpretability, such as renewable energy dispatch, reserve configuration, and operational risk assessment.

[0014] In summary, while existing technologies have made some progress in temporal feature modeling, multi-scale representation, operating condition identification, uncertainty prediction, and semantic enhancement, they have not yet formed an active, hierarchical, and adaptive prediction mechanism for critical risk periods of wind power. This invention addresses this technological gap by proposing a semantically driven, active, hierarchical, and adaptive prediction method for high-risk future intervals. Summary of the Invention

[0015] The technical problem this invention aims to solve is to address the issues of strong volatility, significant non-stationarity, frequent local abrupt events, and concentrated prediction errors in wind power sequences under the background of high proportion of new energy access. It proposes a semantically driven, proactive, hierarchical adaptive wind power prediction method and system capable of actively selecting, locally correcting, and adaptively stopping predictions for high-risk intervals in the future.

[0016] To address the aforementioned technical issues, this invention provides a semantically driven, proactive, hierarchical, adaptive wind power prediction method. This method transforms wind power prediction from a traditional fixed, one-time regression process into a multi-stage dynamic reasoning process under input correlation, risk-driven mechanisms, and semantic constraints. First, an initial prediction and risk characterization of the future power trajectory are obtained through a global prediction module. Then, candidate high-risk intervals are formed through future step-level risk scoring. Subsequently, a high-level semantic context is generated using semantic reprogramming and conditional modeling. In each round of hierarchical reasoning, key prediction intervals, key driving factors, and whether to continue correction are jointly selected. Finally, local incremental correction is performed only on the selected key intervals until the stopping condition is met, at which point the final prediction result is output.

[0017] Preferably, the method includes the following specific steps: Step 1: Acquire and preprocess wind power forecast input data; Step 2: Obtain initial prediction results and internal state through the global prediction module; Step 3: Construct a future step-level risk score and generate a set of candidate high-risk time periods; Step 4: Perform temporal semantic reprogramming and construct cue conditions; Step 5: Jointly determine key forecast intervals, key driving factors, and further correction markers; Step 6: Perform local stepwise correction on the critical interval; Step 7: Update the internal state and perform adaptive stopping judgment; Step 8: Construct joint training objectives and complete end-to-end optimization.

[0018] As a preferred method, acquiring and preprocessing the wind power prediction input data specifically includes: First, acquiring historical multivariate time series data of the wind farm or renewable energy plant to be predicted; represented in the form of historical input sequences as follows: ; in, Indicates the length of the historical input. Indicates the dimension of the input variable. This includes, but is not limited to, historical wind power, wind speed, wind direction, temperature, humidity, atmospheric pressure, numerical weather forecast information, time characteristics, unit operating status, and other site operating information; Represents the set of real numbers. The prediction target is the future. Step wind power sequence: ; The original data is processed in chronological order, including missing value processing, outlier processing, time alignment, and normalization. Missing values ​​are filled using linear interpolation, nearest neighbor interpolation, moving average, or model interpolation. Variables of different dimensions are processed using maximum-minimum normalization or standardization. For data with different sampling frequencies, they are unified to the target sampling interval through resampling or time window aggregation.

[0019] Preferably, in step two, the preprocessed historical input sequence is input into the global prediction module to obtain the initial prediction result and initial internal state of the future power trajectory: ; in, This represents the global prediction module. This represents the initial forecast result for the future forecast interval. This represents the initial risk characterization or internal state; the global prediction module is implemented using convolutional networks, recurrent neural networks, Transformer structures, block-based time series networks, graph neural networks, multilayer perceptrons, or combinations thereof; Step three is specifically implemented as follows: Based on the initial prediction results, construct a future-step-level risk score for each future moment; for each future moment... The risk score is defined as: ; in, This indicates that the global prediction module will be available at future time points. The output or estimated uncertainty measure is used to characterize the degree of lack of confidence in the prediction at that moment; This represents the first-order difference magnitude of the initial predicted trajectory between adjacent time points, used to characterize the intensity of local changes; Represents the prior score of an event, used to measure the time interval. The possibility of critical events such as rapid increase or decrease in wind power, localized switching of operating conditions, or weather changes occurring in the vicinity; These represent the weighting coefficients for uncertainty measure, local change intensity, and prior event score, respectively. It is obtained from the distribution output of the global prediction module, the prediction variance, the ensemble model divergence, the Monte Carlo random inactivation results, or the confidence estimation module; Generated from historical climbing event rules, wind speed change thresholds, sudden wind direction changes, weather state switching, intraday statistical patterns, or an event recognition network obtained through training; Based on risk score and screening threshold The candidate high-risk time set is obtained: ; Subsequently, the sets were arranged according to temporal continuity. The adjacent or nearest neighbor times are aggregated into several candidate key intervals; for isolated noise points, post-processing is performed through minimum duration constraints, neighborhood smoothing, connected segment filtering or interval merging to obtain a set of candidate key intervals.

[0020] Preferably, the input time series is reprogrammed with temporal semantics; first, the input sequence is divided into several time segments: ; in, Indicates the first A time segment, The number of time segments is represented; each time segment is obtained through a sliding window, non-overlapping block, or multi-scale block method; subsequently, a segment encoder encodes each time segment into a latent vector: ; in, This represents a fragment encoder; further, a semantic prototype set is introduced: ; in, Indicates the first Each semantic prototype corresponds to a time series pattern representing stable operation, rapid climb, rapid descent, periodic fluctuation, weather transition, anomalous disturbance, low wind speed range, and high wind speed range. The original numerical fragments are mapped to the semantic prototype space by calculating the association weights between the fragment representations and the semantic prototypes. ; in, Indicates the first The segment and the first The normalized correlation coefficients between the semantic prototypes are used to obtain the temporal semantic representation. ; At the same time, construct the prompt condition vector: ; in, It indicates domain knowledge hints, such as wind power ramp-up events, sudden weather changes, and power curve constraints; This indicates task objective prompts, such as short-term forecasting, ultra-short-term forecasting, multi-step forecasting, and critical interval error suppression. This provides statistical contextual cues, such as mean, variance, trend strength, volatility, recent rate of change, and correlation information. The cues and semantic sequence are input into the context modeling module to obtain a semantic context representation. ; in, It can be implemented using Transformer, cross-attention network, gated fusion network or multilayer perceptron.

[0021] As a preferred option, in the... In successive inference prediction, the model is based on the current internal state. Existing prediction results and semantic context representation The key forecast intervals that need to be prioritized for correction in the current round of joint decision-making Key driving factor set And continue to correct the markings : ; in, Describes the joint decision-making strategy function. This represents the priority correction interval selected from the set of candidate key intervals for the current round. This represents the set of driving variables that require close attention within the current interval. Indicates whether to continue performing local correction; The objective of selecting the key forecast interval is to determine "which future interval should be prioritized for correction in the next step"; the objective of selecting the key driver factors is to determine "which variables dominate the forecast error within that interval"; the key driver factors are implemented through continuous gating vectors. ; in, Indicates the gating weights of the variables. Representing interval context features, and For learnable parameters, The Sigmoid function is used; the reweighted input is obtained based on the gated weights. ; Therefore, the model can dynamically focus on different driving variables in different key intervals. For example, in the stable interval, it can focus on historical power inertia, and in the rapidly fluctuating interval, it can focus on changes in wind speed, wind direction, temperature or air pressure.

[0022] Preferably, when continuing to correct the markings At this time, the model does not re-predict the entire future sequence, but only for the currently selected key interval. Perform local incremental correction; the local correction module outputs the first... Round prediction increment: ; in, This represents a local hierarchical correction module; this module is implemented using convolutional networks, attention networks, residual networks, gated networks, or multilayer perceptrons; to ensure that the correction only applies to the key regions of the target, a region mask is introduced. And update the prediction results: ; in, Only in the key interval Internal activation assigns zero or a small weight to non-critical intervals; through this design, the model can concentrate its computational and correction capabilities on the most error-prone future segments, avoiding indiscriminate repetitive calculations across the entire future prediction interval.

[0023] As a preferred option, complete the first... After local correction, the selected key intervals, key driving factors, corrected prediction results, and semantic context are written back to the internal state: ; in, This indicates the state update module. Subsequently, the benefit assessment value for further correction is calculated: ; in, This represents the profit evaluation function. Based on the stopping threshold... Make a judgment: ; when Or the number of rounds reaches the maximum number of correction rounds. When the time is reached, stop the hierarchical reasoning and output the final prediction result. This mechanism enables the model to perform more rounds of critical interval correction on complex samples and stop early on simple samples, thus balancing prediction accuracy and inference efficiency.

[0024] As a preferred option, to achieve joint optimization of the above modules, the following joint loss function is constructed: ; in, Used to constrain overall prediction accuracy; Used to enhance the model's ability to characterize key event intervals such as climbing, sudden descent, and change of operating conditions; Decision-making strategies for optimizing key intervals and key driving factors; Used to learn adaptive stopping conditions; Used to constrain the consistency and distinctiveness of semantic reprogramming representations; Used to control the additional computational overhead caused by hierarchical correction; to This is a weighting factor.

[0025] Meanwhile, the present invention also provides a semantically driven active hierarchical adaptive wind power prediction system, including a data acquisition and preprocessing module, which is used to acquire historical power, meteorological variables, time characteristics and operation information of wind farms, and perform missing value filling, normalization and time window construction. The global prediction module is used to output the initial prediction results of future power and the initial internal state based on the historical input sequence; The risk scoring module is used to generate future-step-level risk scores based on uncertainty measures, the intensity of local changes, and prior event scores, and to form candidate high-risk intervals. The semantic reprogramming module is used to divide the input time series into segments and map them to the semantic prototype space to obtain a temporal semantic representation. The prompt condition modeling module is used to integrate domain knowledge prompts, task objective prompts, and statistical context prompts to generate semantic context representations; The joint decision-making module is used to select key prediction intervals, key driving factors, and further correction markers based on internal state, existing prediction results, and semantic context representation. The local hierarchical correction module is used to perform local incremental correction on the selected critical intervals; The state update and adaptive stopping module is used to update the internal state, evaluate the benefits of continuing correction, and determine whether to stop hierarchical inference. The results output module is used to output the final wind power prediction results and interpretable information such as key intervals, key factors, and correction cycles.

[0026] After adopting the above method and system, the present invention has the following beneficial effects: First, it improves the prediction accuracy during critical fluctuation periods. The present invention focuses the model's attention on high-risk segments such as uphill climbs, sudden drops, peak-valley transitions, and weather shifts through future-step risk scoring and candidate critical interval generation mechanisms, and reduces prediction bias during critical periods through local hierarchical correction. Comparative experiments published in existing technologies show that on the publicly available Wind dataset, the method of the present invention achieves MAE and MSE of 0.579 and 0.749, respectively, a decrease of 5.4% and 16.0% compared to the suboptimal method; on the self-collected wind farm data_1, the method achieves the optimal MAE of 0.605, a decrease of 5.5% compared to the suboptimal method, and the MSE is basically on par with the strong baseline PatchTST; on the self-collected wind farm data_2, the method achieves MAE and MSE of 0.903 and 0.677, respectively, demonstrating better adaptability to wind power scenarios than existing representative baselines.

[0027] Second, it enhances the model's ability to identify dominant driving factors. Traditional models typically use a uniform modeling approach for all variables, making it difficult to distinguish dominant variables across different future intervals. This invention, through a key factor selection mechanism, dynamically generates variable gating weights in each round of calibration, enabling the model to select different driving factors such as historical power, wind speed, wind direction, temperature, air pressure, or time characteristics based on the evolutionary state of the current key interval, thereby enhancing its ability to express multivariate coupling relationships.

[0028] Third, it enhances the interpretability of the prediction process. This invention not only outputs the final wind power prediction result, but also candidate high-risk intervals, key prediction intervals selected in each round, key driving factors, correction rounds, and reasons for stopping. Compared to black-box models that only output prediction sequences, this invention can explain where the model is corrected, based on which factors, and why correction stops, making it more suitable for business scenarios requiring interpretable evidence, such as power dispatching and operational risk assessment.

[0029] Fourth, it reduces unnecessary computational overhead. Existing methods typically perform uniform calculations across the entire future prediction interval, while this invention uses interval masks to perform local incremental corrections only on high-risk segments and employs an adaptive stopping mechanism to avoid performing too many rounds of corrections on simple samples. Therefore, this invention can reduce redundant calculations for stable periods while maintaining prediction performance during critical periods, thereby improving the efficiency of inference resource utilization.

[0030] Fifth, it enhances generalization ability under semantic conditions. This invention introduces temporal semantic reprogramming and cue condition modeling, mapping the original numerical sequence to a semantic prototype space and integrating domain knowledge, task objectives, and statistical context. This enables the model to make high-level decisions based on "current evolutionary pattern, future risk location, and benefits of further correction," rather than solely relying on the original numerical error for correction. This structure helps improve the model's generalization ability under different wind farms, different meteorological conditions, and different prediction spans.

[0031] Sixth, there is synergistic gain between modules. Ablation experiments in existing technologies show that on the Wind dataset, introducing only the temporal semantic reprogramming module reduces MAE from 2.785 to 1.016 and MSE from 3.579 to 1.159; the complete method further achieves MAE of 0.579 and MSE of 0.749. Existing technologies also indicate that the complete method further reduces MAE by 43.0% and MSE by 35.4% compared to the optimal single-module configuration on the Wind dataset, demonstrating a significant synergistic effect between the semantic reprogramming module and the active hierarchical prediction module. Attached Figure Description

[0032] Figure 1This is the overall network architecture diagram of the semantically driven active hierarchical adaptive wind power prediction method and system of the present invention.

[0033] Figure 2 This is the temporal semantic reprogramming and prompt condition-driven predictive modeling process in this invention.

[0034] Figure 3 This is a comparative experimental result diagram of the present invention and the solutions in the prior art.

[0035] Figure 4 This is the correlation thermodynamic matrix diagram under the initial prediction state in this invention.

[0036] Figure 5 This is a visual comparison chart of the predicted value and the actual value in the initial prediction state of this invention.

[0037] Figure 6 This is the correlation thermodynamic matrix diagram of the final predicted state in this invention.

[0038] Figure 7 This is a visual comparison chart of the predicted value and the actual value in the final predicted state in this invention. Detailed Implementation

[0039] The present invention will now be described in further detail with reference to the accompanying drawings.

[0040] Combined with appendix Figure 1 To the attached Figure 7 This invention provides a semantically driven, proactive, hierarchical, adaptive wind power prediction method and system. This method transforms wind power prediction from a traditional fixed, one-time regression process into a multi-stage dynamic inference process under input correlation, risk-driven mechanisms, and semantic constraints. Its core idea is as follows: First, an initial prediction and risk characterization of the future power trajectory are obtained through a global prediction module. Then, candidate high-risk intervals are formed through future-step-level risk scoring. Subsequently, a high-level semantic context is generated using semantic reprogramming and conditional modeling. In each round of hierarchical inference, key prediction intervals, key driving factors, and whether to continue correction are jointly selected. Finally, local incremental correction is performed only on the selected key intervals until the stopping condition is met, at which point the final prediction result is output.

[0041] Step 1: Acquire and preprocess wind power forecast input data First, obtain historical multivariate time series data of the wind farm or renewable energy plant to be predicted, represented in the form of historical input sequences: ;

[0042] in, Indicates the length of the historical input. Indicates the dimension of the input variable. This includes, but is not limited to, historical wind power, wind speed, wind direction, temperature, humidity, atmospheric pressure, numerical weather forecast information, time characteristics, unit operating status, and other site operating information. Represents the set of real numbers. The prediction target is the future. Step wind power sequence: ;

[0043] The original data is processed in chronological order, including missing value handling, outlier handling, time alignment, and normalization. Missing values ​​are filled using linear interpolation, nearest neighbor interpolation, moving average, or model interpolation. Variables with different dimensions are processed using maximum-minimum normalization or standardization. For data with different sampling frequencies, they are unified to the target sampling interval through resampling or time window aggregation.

[0044] Step 2: Obtain initial prediction results and internal state through the global prediction module. The preprocessed historical input sequence is input into the global prediction module to obtain the initial prediction results and initial internal state of the future power trajectory: ; in, This represents the global prediction module. This represents the initial forecast result for the future forecast interval. This represents the initial risk characterization or internal state. The global prediction module can be implemented using convolutional networks, recurrent neural networks, Transformer structures, block time series networks, graph neural networks, multilayer perceptrons, or combinations thereof.

[0045] In this step, the global prediction module does not directly provide the final high-precision result, but rather establishes a preliminary understanding of future wind power change trends, potential abrupt change locations, and overall risk distribution at a relatively low computational cost. Existing technologies also define this part as establishing an overall understanding of future power change trends and potential risk locations, rather than directly outputting the final result.

[0046] Step 3: Construct a future step-level risk score and generate a candidate high-risk time set. Based on the initial prediction results, a future-step level risk score is constructed for each future moment. For each future moment... The risk score is defined as: ; in, This indicates that the global prediction module will be available at future time points. The output or estimated uncertainty measure is used to characterize the degree of lack of confidence in the prediction at that moment; This represents the first-order difference magnitude of the initial predicted trajectory between adjacent time points, used to characterize the intensity of local changes; Represents the prior score of an event, used to measure the time interval. The possibility of critical events such as rapid increase or decrease in wind power, localized switching of operating conditions, or weather changes occurring in the vicinity; These represent the weighting coefficients for uncertainty measure, local change intensity, and prior event score, respectively.

[0047] It can be obtained from the distribution output of the global prediction module, the prediction variance, the ensemble model divergence, the Monte Carlo random inactivation results, or the confidence estimation module. It can be generated from historical climbing event rules, wind speed change thresholds, sudden wind direction changes, weather state switching, intraday statistical patterns, or trained event recognition networks.

[0048] Based on risk score and screening threshold The candidate high-risk time set is obtained: ; Subsequently, the sets were arranged according to temporal continuity. Adjacent or nearest-neighbor times are aggregated into several candidate key intervals. For isolated noise points, post-processing is performed through minimum duration constraints, neighborhood smoothing, connected segment filtering, or interval merging to obtain a set of candidate key intervals.

[0049] Step 4: Perform temporal semantic reprogramming and build cue conditions To enhance the model's ability to express complex wind power evolution patterns and key driving relationships, this invention performs temporal semantic reprogramming on the input time series. First, the input sequence is divided into several time segments: ; in, Indicates the first A time segment, This indicates the number of segments. Each time segment is obtained using a sliding window, non-overlapping block, or multi-scale block method. Subsequently, a segment encoder encodes each time segment into a latent vector: ; in, This indicates a fragment encoder.

[0050] Further, a semantic prototype set is introduced: ; in, Indicates the first Each semantic prototype corresponds to a typical time series pattern, such as stable operation, rapid climb, rapid descent, periodic fluctuations, weather changes, anomalous disturbances, low wind speed ranges, and high wind speed ranges. By calculating the association weights between fragment representations and semantic prototypes, the original numerical fragments are mapped to the semantic prototype space. ; in, Indicates the first The segment and the first Normalized correlation coefficients between semantic prototypes. This yields the temporal semantic representation: ;

[0051] At the same time, construct the prompt condition vector: ; in, It indicates domain knowledge prompts, such as wind power ramp-up events, sudden weather changes, power curve constraints, etc. This indicates task objectives, such as short-term forecasting, ultra-short-term forecasting, multi-step forecasting, and critical interval error suppression. This represents statistical contextual cues, such as mean, variance, trend strength, volatility, recent rate of change, and correlation information. The cues and semantic sequence are input into the context modeling module to obtain a semantic context representation. ; in, It can be implemented using Transformer, cross-attention network, gated fusion network or multilayer perceptron.

[0052] Step 5: Joint Decision-Making: Key Forecasting Intervals, Key Driving Factors, and Continued Correction Markers In the In successive inference prediction, the model is based on the current internal state. Existing prediction results and semantic context representation The key forecast intervals that need to be prioritized for correction in the current round of joint decision-making Key driving factor set And continue to correct the markings : ; in, Describes the joint decision-making strategy function. This represents the priority correction interval selected from the set of candidate key intervals for the current round. This represents the set of driving variables that require close attention within the current interval. Indicates whether to continue performing local correction.

[0053] The objective of selecting the key forecast interval is to determine "which future interval should be prioritized for correction next"; the objective of selecting the key driver factors is to determine "which variables dominate the forecast error within that interval". In one implementation, the key driver factors are implemented using continuous gating vectors: ; in, Indicates the gating weights of the variables. Representing interval context features, and For learnable parameters, The function is the Sigmoid function. The reweighted input is obtained based on the gated weights: ;

[0054] Therefore, the model can dynamically focus on different driving variables in different key intervals. For example, in the stable interval, it can focus on historical power inertia, and in the rapidly fluctuating interval, it can focus on changes in wind speed, wind direction, temperature or air pressure.

[0055] Step 6: Perform local stepwise correction on the critical intervals. When continuing to correct the markings At this time, the model does not re-predict the entire future sequence, but only for the currently selected key interval. Perform local incremental correction. The local correction module outputs the first... Round prediction increment: ;

[0056] in, This represents a local hierarchical correction module. This module is implemented using convolutional networks, attention networks, residual networks, gating networks, or multilayer perceptrons.

[0057] To ensure that the correction only applies to the critical target region, a region mask is introduced. And update the prediction results: ;

[0058] in, Only in the key interval Internal activation assigns zero or a small weight to non-critical intervals. This design allows the model to focus its computational and correction capabilities on the most error-prone future segments, avoiding indiscriminate and repetitive calculations across the entire future prediction interval.

[0059] Step 7: Update the internal state and perform adaptive stopping judgment. Complete the first After local correction, the selected key intervals, key driving factors, corrected prediction results, and semantic context are written back to the internal state: ;

[0060] in, This indicates the state update module. Subsequently, the benefit assessment value for further correction is calculated: ;

[0061] in, This represents the profit evaluation function. Based on the stopping threshold... Make a judgment: ;

[0062] when Or the number of rounds reaches the maximum number of correction rounds. When the time is reached, stop the hierarchical reasoning and output the final prediction result. This mechanism enables the model to perform more rounds of critical interval correction on complex samples and stop early on simple samples, thus balancing prediction accuracy and inference efficiency.

[0063] Step 8: Construct a joint training objective and complete end-to-end optimization To achieve joint optimization of the above modules, this invention constructs the following joint loss function: ;

[0064] in, Used to constrain overall prediction accuracy; Used to enhance the model's ability to characterize key event intervals such as climbing, sudden descent, and change of operating conditions; Decision-making strategies for optimizing key intervals and key driving factors; Used to learn adaptive stopping conditions; Used to constrain the consistency and distinctiveness of semantic reprogramming representations; Used to control the additional computational overhead caused by hierarchical correction; to This is a weighting factor.

[0065] Corresponding to the above method, the present invention also provides a semantically driven active hierarchical adaptive wind power prediction system, including: a data acquisition and preprocessing module, used to acquire historical power of wind farms, meteorological variables, time characteristics and operation information, and perform missing value filling, normalization and time window construction; The global prediction module is used to output the initial prediction results of future power and the initial internal state based on the historical input sequence; The risk scoring module is used to generate future-step-level risk scores based on uncertainty measures, the intensity of local changes, and prior event scores, and to form candidate high-risk intervals. The semantic reprogramming module is used to divide the input time series into segments and map them to the semantic prototype space to obtain a temporal semantic representation. The prompt condition modeling module is used to integrate domain knowledge prompts, task objective prompts, and statistical context prompts to generate semantic context representations; The joint decision-making module is used to select key prediction intervals, key driving factors, and further correction markers based on internal state, existing prediction results, and semantic context representation. The local hierarchical correction module is used to perform local incremental correction on the selected critical intervals; The state update and adaptive stopping module is used to update the internal state, evaluate the benefits of continuing correction, and determine whether to stop hierarchical inference. The results output module is used to output the final wind power prediction results and interpretable information such as key intervals, key factors, and correction cycles.

[0066] This invention has been experimentally verified using publicly available datasets and self-collected wind farm data. The experiments employed publicly available wind power datasets from Wind, self-collected wind farm datasets, and multivariate time series datasets such as ETT and Electricity. All datasets were divided into training, validation, and test sets in chronological order, and data preprocessing was performed using max-min normalization and interpolation. The experimental environment consisted of Python 3.8 and the PyTorch deep learning framework, running on a Linux system, using an NVIDIA A800 80G GPU, and employing a sliding window method for prediction.

[0067] Example 1: Wind Power Forecasting Based on Wind Public Dataset In this embodiment, the publicly available wind power forecast dataset from Wind was selected as the experimental subject. This dataset contains seven common features for wind power forecasting, including wind speed, wind direction, temperature, humidity, atmospheric pressure, and turbine power, with a sampling interval of 15 minutes. In the experiment, the historical input length was set to 96, and tests were conducted under forecast step sizes of 96, 192, 336, and 720, respectively. The average result across the four forecast lengths was taken as the final evaluation result. This setting can reduce the random influence of a single forecast span on the experimental conclusions.

[0068] The specific implementation process is as follows: First, the Wind dataset is divided into training, validation, and test sets according to time sequence, and missing value interpolation and maximum / minimum value normalization are performed on each variable. Then, a sliding window sample is constructed, with each sample containing a historical input sequence of length 96, and the prediction target is the wind power sequence of a specified future step size.

[0069] Secondly, the historical input sequence is input into the global prediction module to obtain the initial future power prediction results and internal state. Based on the global prediction results, the uncertainty measure, the difference magnitude of the prediction between adjacent times, and the event prior score are calculated for each future time step, and the future step-level risk score is obtained based on the weighted result of the three. Candidate high-risk times are screened according to the risk score threshold, and consecutive or adjacent high-risk times are aggregated into candidate key intervals.

[0070] Next, the historical input sequence is divided into time segments and encoded, and each segment is mapped to a semantic prototype space to form a temporal semantic representation. At the same time, domain knowledge hints, task goal hints, and statistical context hints are constructed and input together with the semantic representation into the context modeling module to obtain a semantic context representation.

[0071] Then, in each round of hierarchical correction, the joint decision-making module selects the most worthy interval for correction from the candidate key intervals based on the current internal state, existing prediction results, and semantic context, and selects the corresponding key driving factor. The local hierarchical correction module reweights the input variables according to the key driving factor and generates prediction increments only on the selected key intervals. The prediction increments are written into the current prediction results through interval masks to obtain the updated prediction sequence.

[0072] Finally, the state update and adaptive stopping module updates the internal state based on the corrected prediction results, the selected key interval, and the key driving factors, and calculates the benefit of continuing correction. When the benefit is lower than a preset threshold or the maximum number of correction rounds is reached, the hierarchical correction stops and the final prediction result is output.

[0073] Experimental results show that on the Wind dataset, the method of this invention achieves an MAE of 0.579 and an MSE of 0.749, representing reductions of 5.4% and 16.0% respectively compared to the suboptimal method. This demonstrates that the present invention can simultaneously improve average error control and large deviation suppression capabilities in public wind power scenarios. (See attached figure.) Figure 4 - Appendix Figure 7 The feature association structure and predicted trajectory under the initial prediction and the stepwise correction states show a significant improvement.

[0074] Example 2: Prediction Verification Based on Self-Collected Wind Farm Data In another embodiment, real wind farm data collected from a wind farm in Guangdong Province between January 1, 2019 and December 31, 2021 was used for verification. This data was collected every 15 minutes and included wind speed and direction at 10m, 30m, and 50m from the wind turbine tower, as well as variables such as temperature, humidity, atmospheric pressure, and turbine power. This region is affected by high temperature and humidity, monsoons, and heavy rain and strong winds, resulting in significant variations in wind speed, direction, temperature, and humidity, making it suitable for verifying the model's predictive ability in real, complex wind power scenarios.

[0075] Following the same process as in Example 1, the self-collected data is first processed for missing values, normalized, and a sliding window is constructed. Then, the global prediction module is used to generate initial prediction results, and the risk scoring module is used to generate candidate key intervals. Next, the semantic reprogramming module and the prompt condition modeling module are used to construct the semantic context. Finally, the key interval-key factor joint decision-making module and the local hierarchical correction module are used to generate the final prediction results.

[0076] Experimental results show that on self-collected wind farm data_1, the method of this invention achieves the optimal MAE of 0.605, which is 5.5% lower than the second-best method, with a corresponding MSE of 0.741, which is basically the same as the 0.739 obtained by PatchTST, with a difference of only 0.26%. On self-collected wind farm data_2, the MAE and MSE of the method of this invention reach 0.903 and 0.677, respectively, which are better than the representative baseline model compared. These results indicate that the present invention is not only applicable to publicly available wind power data, but also maintains good prediction performance and stability in real wind farm data.

[0077] Example 3: Ablation Verification of Module Validity To verify the necessity of each key module in this invention, a baseline model consisting of an encoder and a decoder was constructed, and the performance of the "baseline model," "baseline + temporal semantic reprogramming," "baseline + active hierarchical prediction," and "complete method" were tested respectively. Experimental results show that on the Wind dataset, the baseline model's MAE and MSE are 2.785 and 3.579, respectively; after introducing the temporal semantic reprogramming module, the MAE and MSE decrease to 1.016 and 1.159, respectively; after introducing the active hierarchical prediction module, the MAE and MSE decrease to 1.740 and 1.587, respectively; and the complete method further achieves 0.579 and 0.749.

[0078] The results show that temporal semantic reprogramming can improve the model's ability to express high-level evolutionary patterns of input sequences, and proactive hierarchical prediction can implement targeted corrections around high-risk future intervals. The combination of the two can form synergistic gains, transforming the prediction process from static one-time regression to a dynamic hierarchical correction process oriented towards key time periods.

[0079] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. These computer software products can be stored in computer-readable storage media, such as ROM / RAM, magnetic disks, optical disks, etc., and include several instructions to cause a computer device, such as a personal computer, server, or network device, to execute the methods disclosed in the embodiments.

[0080] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings show only one embodiment of the invention; the actual solution is not limited to this. Besides being applicable to wind farm power prediction, the present invention can also be used for photovoltaic power prediction, source-load net load prediction, new energy grid-connected dispatching, reserve capacity configuration, electricity market clearing, source-grid-load-storage coordinated optimization, distributed energy aggregation and regulation, and other scenarios. It can also be extended to multivariate time series prediction tasks with strong fluctuations, strong non-stationarity, and event-driven characteristics. In conclusion, if those skilled in the art, inspired by this invention, design similar solutions and embodiments without departing from the spirit of the invention, they should all fall within the protection scope of this invention.

Claims

1. A semantic-driven active hierarchical self-adaptive wind power prediction method, characterized in that: The wind power prediction process is transformed from a traditional fixed one-time regression process into a multi-stage dynamic reasoning process under input correlation, risk-driven, and semantic constraints. First, the initial prediction and risk characterization of the future power trajectory are obtained through the global prediction module. Then, candidate high-risk intervals are formed by future step-level risk scoring. Subsequently, high-level semantic context is generated by semantic reprogramming and prompt condition modeling. In each round of hierarchical reasoning, key prediction intervals, key driving factors, and whether to continue correction are jointly selected. Finally, local incremental correction is performed only on the selected key intervals until the stopping condition is met, and the final prediction result is output.

2. The method of claim 1, wherein the method is a semantic-driven, active, hierarchical, self-adaptive wind power prediction method. The specific steps include the following: Step 1: Acquire and preprocess wind power forecast input data; Step 2: Obtain initial prediction results and internal state through the global prediction module; Step 3: Construct a future step-level risk score and generate a set of candidate high-risk time periods; Step 4: Perform temporal semantic reprogramming and construct cue conditions; Step 5: Jointly determine key forecast intervals, key driving factors, and further correction markers; Step 6: Perform local stepwise correction on the critical interval; Step 7: Update the internal state and perform adaptive stopping judgment; Step 8: Construct joint training objectives and complete end-to-end optimization.

3. The semantically driven, proactive, hierarchical adaptive wind power prediction method according to claim 2, characterized in that: The acquisition and preprocessing of wind power prediction input data specifically includes: First, acquiring historical multivariate time series data of the wind farm or renewable energy plant to be predicted; represented in the form of historical input sequences as follows: ， in, Indicates the length of the historical input. Indicates the dimension of the input variable. It displays the station's operational information, including historical wind power, wind speed, wind direction, temperature, humidity, atmospheric pressure, numerical weather forecast information, time characteristics, unit operating status, and other station operational information; Represents the set of real numbers; the prediction target is the future. Step wind power sequence: ， The original data is processed in chronological order, including missing value processing, outlier processing, time alignment, and normalization. Missing values ​​are filled using linear interpolation, nearest neighbor interpolation, moving average, or model interpolation. Variables of different dimensions are processed using maximum-minimum normalization or standardization. For data with different sampling frequencies, they are unified to the target sampling interval through resampling or time window aggregation.

4. The semantically driven, proactive, hierarchical adaptive wind power prediction method according to claim 3, characterized in that: In step two, the preprocessed historical input sequence is input into the global prediction module to obtain the initial prediction results and initial internal state of the future power trajectory: ， in, This represents the global prediction module. This represents the initial forecast result for the future forecast interval. This represents the initial risk characterization or internal state; the global prediction module is implemented using convolutional networks, recurrent neural networks, Transformer structures, block time series networks, graph neural networks, multilayer perceptrons, or a combination of the above tools. Step three is specifically implemented as follows: Based on the initial prediction results, construct a future-step-level risk score for each future moment; for each future moment... The risk score is defined as: ， in, This indicates that the global prediction module will be used in the future. The output or estimated uncertainty measure is used to characterize the degree of lack of confidence in the prediction at that moment; This represents the first-order difference magnitude of the initial predicted trajectory between adjacent time points, used to characterize the intensity of local changes; Represents the prior score of an event, used to measure the time interval. The possibility of critical events such as rapid increase or decrease in wind power, localized switching of operating conditions, or weather changes occurring in the vicinity; These represent the weighting coefficients for uncertainty measure, local change intensity, and prior event score, respectively. It is obtained from the distribution output of the global prediction module, the prediction variance, the ensemble model divergence, the Monte Carlo random inactivation results, or the confidence estimation module; It is generated from historical climbing event rules, wind speed change thresholds, sudden wind direction changes, weather state switching, intraday statistical patterns, or an event recognition network obtained through training. Based on risk score and screening threshold The candidate high-risk time set is obtained: ， Subsequently, the sets were arranged according to temporal continuity. The adjacent or nearest neighbor times are aggregated into several candidate key intervals; for isolated noise points, post-processing is performed through minimum duration constraints, neighborhood smoothing, connected segment filtering or interval merging to obtain a set of candidate key intervals.

5. The semantically driven, proactive, hierarchical adaptive wind power prediction method according to claim 4, characterized in that: Perform temporal semantic reprogramming on the input time series; first, divide the input sequence into several time segments: ， in, Indicates the first A time segment, The number of time segments is represented; each time segment is obtained through a sliding window, non-overlapping block, or multi-scale block method; subsequently, a segment encoder encodes each time segment into a latent vector: ， in, Represents a fragment encoder; introduces a semantic prototype set: ， in, Indicates the first Each semantic prototype corresponds to a time series pattern representing stable operation, rapid climb, rapid descent, periodic fluctuation, weather transition, anomalous disturbance, low wind speed range, and high wind speed range. The original numerical fragments are mapped to the semantic prototype space by calculating the association weights between the fragment representations and the semantic prototypes. ， in, Indicates the first The segment and the first The normalized correlation coefficients between the semantic prototypes are used to obtain the temporal semantic representation. ， At the same time, construct the prompt condition vector: ， in, Indicates domain knowledge hints; This indicates a task objective prompt; This represents a statistical contextual cue; the cue conditions and semantic sequence are input into the context modeling module to obtain a semantic contextual representation: ， in, It can be implemented using Transformer, cross-attention network, gated fusion network or multilayer perceptron.

6. The semantically driven active hierarchical adaptive wind power prediction method according to claim 5, characterized in that: In the In successive inference prediction, the model is based on the current internal state. Existing prediction results and semantic context representation The key forecast intervals that need to be prioritized for correction in the current round of joint decision-making Key driving factor set And continue to correct the markings : ， in, Describes the joint decision-making strategy function. This represents the priority correction interval selected from the set of candidate key intervals for the current round. This represents the set of driving variables that require close attention within the current interval. Indicates whether to continue performing local correction; The goal of selecting the key forecast interval is to determine which future interval should be prioritized for correction in the next step; the goal of selecting the key driver factors is to determine which variables dominate the forecast error within that interval; the key driver factors are implemented through continuous gating vectors. ， in, Indicates the gating weights of the variables. Representing interval context features, and For learnable parameters, The Sigmoid function is used; the reweighted input is obtained based on the gated weights. ， Therefore, the model dynamically focuses on different driving variables in different key intervals.

7. The semantically driven, proactive, hierarchical adaptive wind power prediction method according to claim 6, characterized in that: When continuing to correct the markings At this time, the model does not re-predict the entire future sequence, but only for the currently selected key interval. Perform local incremental correction; the local correction module outputs the first... Round prediction increment: ， in, This represents a local hierarchical correction module; this module is implemented using convolutional networks, attention networks, residual networks, gated networks, or multilayer perceptrons; a region mask is introduced. And update the prediction results: ， in, Only in the key interval Internal activation, taking zero or a small weight in non-critical intervals.

8. The semantically driven active hierarchical adaptive wind power prediction method according to claim 7, characterized in that: Complete the first After local correction, the selected key intervals, key driving factors, corrected prediction results, and semantic context are written back to the internal state: ， in, This indicates the state update module, which then calculates the estimated benefit of further correction: ， in, This represents the return evaluation function, based on the stopping threshold. Make a judgment: ， when Or the number of cycles reaches the maximum number of correction cycles. When the time is reached, stop the hierarchical reasoning and output the final prediction result. .

9. The semantically driven active hierarchical adaptive wind power prediction method according to claim 8, characterized in that: To achieve joint optimization of the above modules, the following joint loss function is constructed: ， in, Used to constrain overall prediction accuracy; Used to enhance the model's ability to characterize key event intervals such as climbing, sudden descent, and change of operating conditions; Decision-making strategies for optimizing key intervals and key driving factors; Used to learn adaptive stopping conditions; Used to constrain the consistency and distinctiveness of semantic reprogramming representations; Used to control the additional computational overhead caused by hierarchical correction; to This is a weighting factor.

10. A semantically driven, proactive, hierarchical, adaptive wind power prediction system, characterized in that: It includes a data acquisition and preprocessing module, which is used to acquire historical power, meteorological variables, time characteristics and operation information of wind farms, and perform missing value filling, normalization and time window construction; The global prediction module is used to output the initial prediction results of future power and the initial internal state based on the historical input sequence; The risk scoring module is used to generate future-step-level risk scores based on uncertainty measures, the intensity of local changes, and prior event scores, and to form candidate high-risk intervals. The semantic reprogramming module is used to divide the input time series into segments and map them to the semantic prototype space to obtain a temporal semantic representation. The prompt condition modeling module is used to integrate domain knowledge prompts, task objective prompts, and statistical context prompts to generate semantic context representations; The joint decision-making module is used to select key prediction intervals, key driving factors, and further correction markers based on internal state, existing prediction results, and semantic context representation. The local hierarchical correction module is used to perform local incremental correction on the selected critical intervals; The state update and adaptive stopping module is used to update the internal state, evaluate the benefits of continuing correction, and determine whether to stop hierarchical inference. The results output module is used to output the final wind power prediction results and interpretable information such as key intervals, key factors, and correction cycles.

Images