Partition optimization scheduling control method and system for energy storage power station

Abstract

The invention discloses a partition optimization scheduling control method and system for an energy storage power station, and relates to the technical field of power grid load prediction scheduling, and the method comprises the following steps: constructing a multi-scale dynamic baseline of load data, carrying out the phase calibration of a short-time fluctuation signal and a long-term trend signal under a unified time coordinate, and obtaining a multi-scale dynamic baseline; generating a cross-scale coupling mapping matrix by using a time embedding operator based on a calibration result; and establishing a gradient phase tracking network based on the cross-scale coupling mapping matrix. According to the method, time alignment and self-adaptive correction of short-time fluctuation and long-term trend signals are achieved through a multi-scale dynamic baseline and gradient phase tracking mechanism, weight energy balance is kept, and prediction continuity and accuracy are remarkably improved. Meanwhile, a causal residual error playback chain and time scale self-folding mapping enable the model to have self-correction and self-stabilization capabilities, sub-models can be automatically split during phase resonance, multi-scale response coordination is ensured, and the stability and operation efficiency of power grid load prediction scheduling are improved.

Description

Technical Field

[0001] This invention relates to the field of power grid load forecasting and dispatching technology, specifically to a zoned optimization dispatching control method and system for energy storage power stations. Background Technology

[0002] Zonal optimization dispatch control of energy storage power stations refers to the use of artificial intelligence, big data analysis, and predictive algorithms in a multi-regional power grid operation environment to model and predict future power load trends and renewable energy output in each zone. Combined with the regional power grid operation characteristics, this involves dynamic coordination and optimal control of the charging and discharging behavior, energy allocation ratio, and operating sequence of energy storage power stations. The core idea is to divide energy storage power stations into multiple functional zones based on geographical location, grid topology, and load characteristics within a unified dispatch framework. By predicting future power supply and demand differences in each zone, cross-regional coordination and local autonomy of energy storage resources are achieved. Based on the prediction results and real-time operating data, the dispatch control system intelligently determines the charging and discharging timing and power allocation of energy storage units in each zone, achieving a dual balance of power in both the temporal and spatial domains. This facilitates peak shaving and valley filling, enhances renewable energy absorption capacity, and strengthens the grid's resilience and economic operation when dealing with sudden load changes or renewable energy fluctuations.

[0003] The existing technology has the following shortcomings:

[0004] In the zonal optimization dispatch control process of energy storage power stations, real-time load forecasting chains typically need to process both short-term fluctuation data and long-term trend data simultaneously to ensure the continuity and foresight of dispatch decisions. However, when the power system enters a phase of drastic load changes (such as a sudden surge in peak load or a rapid increase in the proportion of renewable energy connected to the grid), the random drift of short-term fluctuation data and the slow drift of long-term trend data may create a coupled resonance effect within the model. Because the two types of drifts differ significantly in time scale and phase characteristics, their superposition can trigger unstable evolution of the parameter gradients within the prediction model, leading to nonlinear collapse of the weight matrix during adaptive updates. This collapse causes the model's prediction distribution to lose its original structure in a very short time, degrading the output characteristics to the level of random noise. Consequently, the intelligent dispatching system executes invalid or reverse dispatching commands based on distorted predictions, potentially leading to severe consequences such as regional power grid frequency imbalance, increased voltage fluctuations, and even dispatch link collapse. Summary of the Invention

[0005] The purpose of this invention is to provide a zoned optimization scheduling and control method and system for energy storage power stations to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a zoned optimization scheduling and control method for an energy storage power station, comprising the following steps:

[0007] Step 1: Construct a multi-scale dynamic baseline for load data, perform phase calibration on short-term fluctuation signals and long-term trend signals under a unified time coordinate, and generate a cross-scale coupling mapping matrix based on the calibration results using a time embedding operator to establish a continuous correspondence between features at multiple time scales, providing a continuous time reference for subsequent gradient dynamic constraints.

[0008] Step 2: Establish a gradient phase tracking network based on the cross-scale coupling mapping matrix, continuously analyze the phase difference between short-term fluctuation signals and long-term trend signals, extract phase drift gradient information, and calculate and generate weight stability coefficients based on the phase drift gradient to form a dynamic constraint basis for the gradient update rate within the constraint model.

[0009] Step 3: Based on the weight stability coefficient, an adaptive weight migration mechanism is constructed to perform energy migration and redistribution operations on the weight channels in the gradient phase tracking network that have drift anomalies, forming a multi-scale weight balance structure to maintain the consistency and dynamic balance of weight energy during parameter update.

[0010] Step 4: Based on the output of the multi-scale weight balance structure, establish a causal residual playback chain, perform time-series reverse mapping and paired write-back on the residuals output by the power grid load prediction model, and dynamically close-loop correct the parameters of the gradient phase tracking network according to the correlation between historical drift and real-time deviation, thereby realizing self-feedback correction of prediction error during training.

[0011] Step 5: Based on the stable operation of the causal residual playback chain, a time-scale self-folding mapping is introduced to project short-term fluctuation signals and long-term trend signals to different time and frequency domains, respectively, to construct a fast-response prediction kernel and a slow-change trend kernel. The phase conjugate operator is used to perform energy synchronization correction on the prediction kernel, so that the power grid load prediction model automatically splits into independent sub-models when phase resonance is detected, thereby preventing gradient interference between signals of different time scales from the source and maintaining the dynamic stability of the prediction model.

[0012] Preferably, the step of constructing a multi-scale dynamic baseline for load data specifically includes the following sub-steps:

[0013] The original power grid load data is decomposed into multiple scales and mapped to a unified time. Key elements such as active load, voltage, current and meteorological conditions are extracted from historically collected power grid operation data. By sampling with a unified time resolution and layering of time intervals, the complete load sequence is divided into short-term fluctuation signal sequence and long-term trend signal sequence. The two types of signals are then corrected by time scale to achieve complete alignment of the time dimension.

[0014] After establishing a unified time coordinate, the phase of short-term fluctuation signals and long-term trend signals is calibrated. The inflection point is determined by calculating the direction and rate of change of the two types of signals. The phase difference is smoothed within the sliding time window to form a continuous time mapping sequence, so as to establish a phase alignment reference covering the entire time range.

[0015] After obtaining the phase calibration results, a cross-scale coupling mapping matrix is ​​generated using the time embedding operator. The phase correspondence is used as the time embedding input. A set of time nodes is set on a unified time axis and a continuous interpolation relationship is established to ensure that signals at different time scales maintain consistency in phase, amplitude and energy distribution, thereby forming a continuous time reference structure that spans the entire time domain.

[0016] Preferably, the step of calculating the weight stability coefficients based on the phase drift gradient includes the following sub-steps:

[0017] The generated cross-scale coupling mapping matrix is ​​expanded in the time domain and the phase difference is extracted. The phase correspondence information between the short-term fluctuation signal and the long-term trend signal at each moment is extracted, and the phase change direction and rate of change between adjacent time points are identified to form the phase drift curve of the short-term fluctuation signal relative to the long-term trend signal.

[0018] After obtaining continuous phase difference data, the changing characteristics of the phase drift curve are analyzed and identified to determine the duration and amplitude of the rising, falling and stable intervals, forming a phase drift time correlation sequence that reflects the interaction characteristics of short-term fluctuation signals and long-term trend signals.

[0019] Based on the phase drift time correlation sequence, phase drift gradient information is extracted, and a gradient change reference line is established according to the continuous time change trend. The phase drift direction and persistence are determined, forming a dynamic gradient input that characterizes the stability of phase coupling.

[0020] The weight stability coefficient is calculated based on the extracted phase drift gradient information. The phase stability of the system is determined by analyzing the amplitude change and direction switching frequency of the phase drift gradient. A higher weight stability coefficient is generated when the drift gradient is gentle and the direction is consistent, and a lower weight stability coefficient is generated when the drift gradient fluctuates violently, so as to achieve dynamic constraint and balance adjustment of the gradient update rate inside the model.

[0021] Preferably, in the step of extracting the phase, the consistency of direction and rate of change of the phase drift time correlation sequence within a continuous time interval are comprehensively judged. Only when the phase drift direction remains continuous and the change amplitude is stable within a preset time threshold is the drift amount of the continuous time interval included in the effective gradient range, so as to ensure that the extracted phase drift gradient information has continuity in time and physical stability.

[0022] Preferably, the step of performing energy transfer and redistribution operations on weight channels exhibiting drift anomalies in the gradient phase tracking network to form a multi-scale weight equilibrium structure includes the following sub-steps:

[0023] Based on the weight stability coefficient, the energy state of each weight channel in the gradient phase tracking network is continuously monitored and its features are identified. By comparing the rate of change and relative amplitude of channel parameters at different times, channels with abnormal energy distribution are identified and an overall energy distribution map is formed.

[0024] After identifying the weighted channel with abnormal energy distribution, an energy migration operation is performed. The neighboring channel with high time correlation and strong phase stability with the energy overload channel is selected as the energy receiving direction. The energy is dynamically transferred between channels by smoothly adjusting the rate of change of parameters, and the migration rate and amplitude are limited by the weight stability coefficient to maintain time continuity.

[0025] After completing the energy transfer, an energy redistribution operation is performed on each weighted channel. High-frequency channels and low-frequency channels are distinguished according to the time scale characteristics of the corresponding channels. The energy ratio of the channels is adjusted according to the energy distribution state, so that short-term fluctuation signals and long-term trend signals can be synchronously expressed in the weight space and maintain dynamic energy balance.

[0026] Preferably, in the energy redistribution operation, the energy adjustment range of the high-frequency channel and the low-frequency channel is dynamically constrained by the weight stability coefficient. When the weight stability coefficient is higher than the set threshold, a small energy adjustment is performed to maintain channel balance. When the weight stability coefficient is lower than the set threshold, the energy migration ratio is increased to accelerate the restoration of equilibrium, thereby ensuring that the multi-scale weight balance structure maintains the stability and continuity of energy distribution during the time evolution process.

[0027] Preferably, the step of establishing a causal residual playback chain based on the output of a multi-scale weighted equilibrium structure includes the following sub-steps:

[0028] Based on the output of the multi-scale weighted equilibrium structure, the residual between the model prediction and the actual load observation is obtained, and the residual is decomposed into short-term fluctuation components and long-term trend components and time-aligned to form a residual time series with cross-scale correspondence.

[0029] After obtaining the residual time series, a time-series reverse mapping path for the residuals is established to trace the prediction error backward along the time axis to its formation stage, and the causal source of the residuals is determined based on the phase correspondence between the short-term fluctuation signal and the long-term trend signal.

[0030] After identifying the causal source of the residuals, the residuals are paired and written back, applied to the corresponding high-frequency weight channel or low-frequency weight channel, and the amplitude and rate of the residual writing back are dynamically constrained based on the weight stability coefficient.

[0031] After completing the residual pairing and write-back, the temporal continuity of the causal residual playback chain is maintained and closed-loop correction is performed. By periodically monitoring the residual change trend and write-back effect, a time-series self-stabilizing closed-loop structure is formed between model output, error detection and parameter correction.

[0032] Preferably, the step of introducing time-scale self-folding mapping based on the stable operation of the causal residual playback chain includes the following sub-steps:

[0033] Based on the stable operation of the causal residual playback chain, the time scale layering and self-folding mapping of the time output sequence of the multi-scale weighted equilibrium structure are performed, folding the original time axis into short time scale segments and long time scale segments, while maintaining phase continuity and energy conservation at the mapping boundary.

[0034] The short-time fluctuation signal and the long-term trend signal after self-folding mapping are projected to different time and frequency domains, so that the short-time fluctuation signal is projected to the high-frequency domain and the long-term trend signal is projected to the low-frequency domain, thereby achieving time and frequency separation of energy.

[0035] After separating the high and low frequency domains, fast response prediction kernels and slow change trend kernels are constructed respectively, and energy transfer nodes are established at the boundary points of the time folding mapping to maintain continuous energy transition.

[0036] A phase conjugate operator is introduced to synchronously correct the energy distribution and phase relationship of the high- and low-frequency prediction kernels, making them phase complementary at the boundary to eliminate the phase resonance phenomenon.

[0037] When a sustained phase resonance or enhanced energy coupling is detected between high and low frequency signals in the time-frequency domain, the power grid load prediction model structure is automatically split, decomposing the overall model into independent sub-models to maintain the dynamic stability of the prediction process.

[0038] A zoned optimization scheduling and control system for an energy storage power station includes a multi-scale dynamic baseline construction module, a gradient phase tracking module, an adaptive weight transfer module, a causal residual playback module, and a time-scale self-folding module.

[0039] The multi-scale dynamic baseline construction module constructs a multi-scale dynamic baseline for load data, performs phase calibration on short-term fluctuation signals and long-term trend signals under a unified time coordinate, and generates a cross-scale coupling mapping matrix based on the calibration results using a time embedding operator.

[0040] The gradient phase tracking module establishes a gradient phase tracking network based on a cross-scale coupling mapping matrix, continuously analyzes the phase difference between short-term fluctuation signals and long-term trend signals, extracts phase drift gradient information, and calculates and generates weight stability coefficients based on the phase drift gradient.

[0041] The adaptive weight transfer module builds an adaptive weight transfer mechanism based on the weight stability coefficient. It performs energy transfer and redistribution operations on weight channels that have drift anomalies in the gradient phase tracking network, forming a multi-scale weight balance structure.

[0042] The causal residual playback module establishes a causal residual playback chain based on the output results of the multi-scale weight balance structure. It performs time-series reverse mapping and paired write-back on the residuals output by the power grid load prediction model, and dynamically closes the loop to correct the parameters of the gradient phase tracking network according to the correlation between historical drift and real-time deviation.

[0043] The time-scale self-folding module, based on the stable operation of the causal residual playback chain, introduces time-scale self-folding mapping to project short-term fluctuation signals and long-term trend signals to different time and frequency domains respectively, constructing a fast-response prediction kernel and a slow-change trend kernel, and using the phase conjugate operator to perform energy synchronization correction on the prediction kernel, so that the power grid load prediction model automatically splits into independent sub-models when phase resonance is detected.

[0044] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0045] This invention introduces a multi-scale dynamic baseline and gradient phase tracking mechanism into the load forecasting process, enabling the model to achieve continuous alignment in the time dimension and adaptive phase difference correction when facing the interaction of short-term fluctuation signals and long-term trend signals. This process establishes a stable phase mapping relationship between short-term and long-term features under a unified time coordinate. During the model update process, the gradient step size is automatically adjusted according to the weight stability coefficient, avoiding gradient oscillations and parameter collapse caused by multi-timescale drift. By maintaining the dynamic balance of weight energy, the model can maintain structural stability and output consistency during the operational phase where rapid load fluctuations and gradual trend changes occur alternately, thereby significantly improving the temporal continuity and dynamic accuracy of load forecasting.

[0046] This invention constructs a causal residual playback chain and a time-scale self-folding mapping mechanism, enabling the model to possess closed-loop self-correction capabilities for historical biases and real-time errors. Furthermore, it actively triggers sub-model splitting upon detecting phase resonance between multi-scale signals, achieving adaptive steady-state adjustment at the structural level. This mechanism allows the model to identify the sources of prediction biases in real-time and perform energy synchronization correction in complex power grid operating environments, ensuring that prediction responses at different time scales maintain independent and coordinated evolutionary characteristics. Through this time-series closed-loop and structural self-separation approach, the model possesses continuous self-stability and anti-interference capabilities during long-term operation, effectively reducing the risk of scheduling link instability and improving the stability and overall operational efficiency of power grid load forecasting and scheduling. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0048] Figure 1 This is a flowchart of a zoned optimization scheduling control method for an energy storage power station according to the present invention;

[0049] Figure 2 A schematic diagram illustrating the principle of constructing a multi-scale dynamic baseline for load data in this invention;

[0050] Figure 3 This is a schematic diagram illustrating the principle of introducing time-scale self-folding mapping based on the stable operation of the causal residual playback chain in this invention.

[0051] Figure 4 This is a schematic diagram of a partitioned optimized scheduling and control system for an energy storage power station according to the present invention. Detailed Implementation

[0052] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0053] This invention provides, for example Figures 1 to 3 The illustrated method for zoned optimization scheduling and control of an energy storage power station includes the following steps:

[0054] Step 1: Construct a multi-scale dynamic baseline for load data, perform phase calibration on short-term fluctuation signals and long-term trend signals under a unified time coordinate, and generate a cross-scale coupling mapping matrix based on the calibration results using a time embedding operator to establish a continuous correspondence between features at multiple time scales, providing a continuous time reference for subsequent gradient dynamic constraints.

[0055] Based on the calibration results, a cross-scale coupling mapping matrix is ​​generated using a time embedding operator. The specific implementation steps are as follows:

[0056] The original power grid load data undergoes multi-scale decomposition and unified temporal mapping. Specifically, key elements such as active load, voltage, current, and meteorological conditions for continuous time periods are first extracted from historically collected power grid operation data and sampled at a unified time resolution, ensuring temporal alignment of data from different sources. Subsequently, the complete load sequence is divided into short-term fluctuation signal sequences and long-term trend signal sequences through time interval layering. Short-term fluctuation signal sequences primarily reflect rapid load changes at the minute or hour level, such as load disturbances caused by sudden weather changes, concentrated user power consumption, or instantaneous fluctuations in renewable energy output. Long-term trend signal sequences reflect daily, weekly, monthly, and even seasonal variations, exhibiting the overall trend of gradual load evolution over time. To ensure the comparability of short-term fluctuation signals and long-term trend signals on the same time axis, time-scale correction is performed on both types of signals. The sampling interval of the short-term fluctuation signal is mapped to an integer multiple of the long-term trend signal time interval through interpolation expansion or resampling, thereby achieving complete temporal alignment. Subsequently, to ensure the continuity and traceability of the time coordinates, a unified reference point based on the actual operating time is established on the time axis. The load state at each moment is defined as a superposition of short-term fluctuation signals and long-term trend signals, so that the two types of signals have a unified time identifier and energy reference at the same point in time. Through this process, the original multi-source asynchronous data is transformed into a continuous and uniform time series, ensuring that the short-term fluctuation signals and long-term trend signals can be characterized and dynamically analyzed under the same time baseline in subsequent steps.

[0057] After establishing a unified time coordinate, phase calibration is performed on short-term fluctuation signals and long-term trend signals. The key to this process is identifying and aligning the characteristic change points of the two types of signals in the same time domain, thereby revealing their temporal correspondence and phase difference distribution. Specifically, each sampling point in the unified time series is first traversed, and the direction and rate of change of the short-term fluctuation signal and the long-term trend signal at that moment are calculated. The inflection point is determined by comparing the positive and negative trends of load changes at adjacent time points. The time position at the inflection point is considered a key phase node of the signal. After identifying the local peaks, valleys, and inflection points of all short-term fluctuation signals, these key nodes are compared one by one with the characteristic nodes of the long-term trend signal, and the offset between them in the same time coordinate is recorded. Subsequently, this offset is smoothed to form a continuous time mapping sequence describing the phase difference change between the short-term fluctuation signal and the long-term trend signal. In this process, to prevent short-term abnormal disturbances from affecting the overall calibration results, a sliding time window mechanism is introduced, updating the overall phase correspondence only when the phase difference is continuously stable within the window. This processing allows for the precise acquisition of the phase position and direction of change of the short-term fluctuation signal relative to the long-term trend signal at any given time point, clearly characterizing the dynamic relationship between the two types of signals on the time axis. Finally, by performing time-weighted integration on the phase-corresponding sequences, a phase alignment reference covering the entire time range is established, ensuring that local changes in the short-term fluctuation signal remain consistent with the overall evolution of the long-term trend signal. This provides a high-precision phase registration basis for subsequent time embedding processes.

[0058] After obtaining the phase calibration results, a cross-scale coupling mapping matrix is ​​generated using a time embedding operator to form a continuous time reference structure spanning the entire time domain. In the specific implementation, the phase correspondence obtained in the previous step is used as the core input for time embedding, and the relative phase, time position, and direction of change of the short-term fluctuation signal and the long-term trend signal at each moment are structurally correlated. To this end, a set of time nodes with fixed intervals is set on a unified time axis, with each node simultaneously corresponding to the short-term fluctuation signal value, the long-term trend signal value, and their phase difference state. A continuous interpolation relationship is established between these nodes, ensuring that signals at different time scales have smooth transition characteristics within each time interval. Subsequently, by calculating the phase change trend between adjacent nodes moment by moment, this continuous phase correspondence is mapped into a set of time-pair mappings describing multi-scale time dependence. Based on this, the time embedding operator is applied to the above mapping set, so that the time state at each moment not only includes the current signal value but also carries the phase transfer information and trend evolution direction between previous and subsequent moments. The generated cross-scale coupling mapping matrix can fully describe the dynamic coupling relationship between short-term fluctuation signals and long-term trend signals in the time domain, and maintain the consistency of the correspondence between the two types of signals in phase, amplitude, and energy distribution in a continuous-time form. To ensure the stability and reliability of this mapping matrix in subsequent use, the continuity of phase increments is maintained during the generation process to avoid abrupt jumps or discontinuous intervals, thereby ensuring that the entire time reference structure is smooth, differentiable, and physically interpretable in the time dimension. Through the above process, short-term fluctuation signals and long-term trend signals form an organic whole under the same time coordinate, enabling a stable and traceable temporal correspondence between cross-scale features. The generated cross-scale coupling mapping matrix not only reflects the real physical laws of grid load changes over time, but also provides a unified, continuous, and self-consistent time baseline for subsequent gradient dynamic constraints and parameter updates, enabling the prediction model to maintain stability and predictability in complex scenarios where changes at different time scales coexist.

[0059] Through the above steps, short-term fluctuation signals and long-term trend signals are gradually transformed from their original separate state into a dynamic, integrated expression within a unified time frame, realizing a complete process from data scale fusion to phase coupling and then to the establishment of a time reference. The resulting multi-scale dynamic baseline not only eliminates structural differences and phase mismatches between signals of different time scales but also forms a continuous time reference system that can be directly referenced by subsequent gradient dynamic constraints.

[0060] Step 2: Establish a gradient phase tracking network based on the cross-scale coupling mapping matrix, continuously analyze the phase difference between short-term fluctuation signals and long-term trend signals, extract phase drift gradient information, and calculate and generate weight stability coefficients based on the phase drift gradient to form a dynamic constraint basis for the gradient update rate within the constraint model.

[0061] The weight stability coefficients are calculated based on the phase drift gradient. The specific implementation steps are as follows:

[0062] First, the generated cross-scale coupling mapping matrix is ​​expanded in the time domain and its phase difference is extracted. In the specific implementation, the mapping matrix is ​​read continuously in chronological order, extracting the phase correspondence information between the short-term fluctuation signal and the long-term trend signal at each moment. Since the mapping matrix has achieved time and phase synchronization during the construction phase, the relative phase positions of the two types of signals on the same time axis can be directly obtained during the analysis process. By traversing the mapping matrix moment by moment, the direction and rate of phase change between adjacent time points are identified, thus forming the phase drift curve of the short-term fluctuation signal relative to the long-term trend signal. To avoid phase jumps caused by sudden load changes or noise interference, the phase drift curve is continuously corrected to maintain a smooth and progressive trend over time. After this processing, each time point corresponds to a set of phase difference information with clear physical meaning, providing a stable input for subsequent extraction of phase drift features.

[0063] After obtaining continuous phase difference data, the phase drift relationship between short-term fluctuation signals and long-term trend signals is analyzed and its features are identified. Specifically, the direction of change of the phase drift curve is tracked along the time axis, and its rising, falling, and stable interval characteristics in different time intervals are analyzed. The rising interval indicates that the phase of the short-term fluctuation signal gradually advances relative to the long-term trend signal, indicating that short-cycle load fluctuations in the system have a leading effect on the overall trend; the falling interval indicates that the phase of the short-term fluctuation signal lags behind the long-term trend signal, reflecting the absorption and delay effect of long-term load evolution on short-term disturbances; the stable interval means that the phases of short-term fluctuations and long-term trends are synchronized, and the system is in a dynamic equilibrium state. During the analysis process, the duration, amplitude, and temporal distribution of phase drift changes in these different intervals are further recorded to form a temporal correlation sequence of phase drift. This sequence reflects the degree of mutual influence between the short-term fluctuation signal and the long-term trend signal at different operating stages and reveals the energy transfer path across time scales during load changes. Through this process, the dynamic evolution trajectory of phase drift in the time domain can be completely depicted, providing accurate time references and phase correlation basis for generating gradient dynamic constraints.

[0064] After obtaining the time-correlated sequence of phase drift, the phase drift gradient information is extracted, and a gradient change reference line is established based on the continuous time-varying trend. Specifically, the rate and direction of phase drift change are calculated along the time axis, and the phase difference between adjacent time points is analyzed as the local drift amount. By accumulating the drift amount over a continuous time period, the strength distribution and directional consistency of the phase drift can be identified. When short-term fluctuation signals are continuously advanced within a certain time period, the drift gradient shows a positive cumulative trend, indicating that the short-term characteristics of the system have an enhanced influence on the long-term trend; when short-term fluctuation signals are continuously lagging, the drift gradient shows a negative distribution, indicating that the overall system trend has an absorption and delay effect on short-term disturbances. To prevent local over-response caused by extreme events or abnormal disturbances, a continuity judgment condition is introduced when extracting the drift gradient. Only when the phase drift maintains a consistent direction over time and the duration exceeds a set threshold can the change be included in the effective gradient range. The resulting phase drift gradient information not only reflects the direction of mutual influence between short-term and long-term signals in different time periods but also reflects the stability of phase coupling at the intensity level. This process effectively combines continuous time variation with phase energy transfer, providing precise dynamic input for the next step of generating weighted stability coefficients.

[0065] Finally, based on the extracted phase drift gradient information, weight stability coefficients are calculated and generated to constrain the rate and direction of gradient updates within the model. In practice, the phase drift gradient is treated as a continuous time signal and accumulated. By observing its amplitude changes and direction switching frequency over time, the phase stability of the system within that time period is determined. When the phase drift gradient changes smoothly and in a consistent direction, it indicates that the short-term fluctuation signal and the long-term trend signal are in a relatively stable coupling state. In this case, the generated weight stability coefficient should be relatively high to allow the model to maintain a normal learning rate during this stage. When the phase drift gradient exhibits violent fluctuations or frequent reversals, it indicates that the coupling relationship between short-term fluctuations and the long-term trend is undergoing an unstable phase. In this case, a lower weight stability coefficient needs to be generated to suppress violent gradient updates within the model, thereby avoiding unnecessary oscillations or collapses of the weights over time. By continuously calculating the weight stability coefficients on the time axis, the model's gradient update process possesses time-adaptive capabilities, automatically adjusting the update intensity and direction under different phase stability conditions to maintain the balance and consistency of parameter evolution. The introduction of this weight stability coefficient essentially provides a dynamic constraint reference for the subsequent adaptive weight transfer, enabling the entire gradient phase tracking process to form a closed-loop adjustment relationship, ensuring that the model can maintain stable, coordinated and physically consistent dynamic response characteristics under the interaction of signals across time scales.

[0066] The specific method for determining that the phase drift gradient change is gradual is as follows:

[0067] Statistical analysis of the rate of change of the phase drift gradient within a continuous time window is performed. If the gradient change amplitude between adjacent time points remains within a set threshold range, and the gradient change direction remains consistent or changes slowly at most sampling points, the phase drift gradient is considered to be changing smoothly. In this case, the gradient standard deviation within the time window is small, and the number of direction reversals is less than a preset proportion of the total number of samples (e.g., below 10%), indicating that the phase difference between the short-term fluctuation signal and the long-term trend signal is in a continuous and consistent evolutionary state, and the energy transfer process is stable without sudden disturbances. Under these circumstances, the update of the model's internal parameters can be regarded as a linear response, the continuity of the predicted features is good, and the gradient evolution can remain smooth and predictable over time.

[0068] The specific method for determining that the phase drift gradient has experienced drastic fluctuations is as follows:

[0069] Within the same time window, the rate of change and direction switching frequency of the phase drift gradient are calculated. If the gradient change between adjacent sampling points exceeds a set threshold multiple times, and the number of direction reversals exceeds a preset proportion of the total number of samples (e.g., more than 30%), it is determined that the gradient fluctuates violently. At this time, the variance of the gradient increases significantly, and the positive and negative drift directions alternate frequently, indicating that the coupling state between the short-term fluctuation signal and the long-term trend signal is unstable, and there is a phase resonance or energy cross-interference phenomenon. The internal parameter update of the model in this stage exhibits nonlinear perturbation characteristics, which may lead to weight energy concentration, update oscillation, or even local collapse. Therefore, it is necessary to immediately reduce the update rate or adjust the weight constraint strength to suppress gradient divergence and restore the dynamic stability of the model.

[0070] It should be noted that:

[0071] The calculation of the weight stability coefficient is based on the phase drift gradient change characteristics between short-term fluctuation signals and long-term trend signals. By comprehensively evaluating the dynamic rate of change, persistence, and energy balance of the phase difference between the two types of signals under a unified time coordinate, the stability of model weight updates within each time period is determined, thus forming a dynamic factor for constrained gradient adjustment. Specifically, in the continuous time series, the phase difference change between the short-term fluctuation signal and the long-term trend signal at adjacent time points is first calculated, and the average absolute rate of change is calculated within a sliding time window to reflect the overall intensity of the phase drift. Then, the consistency ratio of the phase drift direction within this time window is statistically analyzed. If the phase change direction remains consistent at most time points, it indicates that the system is in a relatively stable state; conversely, if the direction frequently reverses, it indicates that the system has a tendency for phase oscillation or resonance. Furthermore, the difference in energy distribution between the short-term and long-term signals within this time window is compared, for example, by calculating the standardized difference of the amplitudes of the two types of signals to measure the degree of energy matching. Finally, the three quantitative indicators (phase change rate, consistency ratio, and energy difference) are weighted and fused to generate a weighted stability coefficient between 0 and 1. The closer the value is to 1, the more stable the model state is, and the gradient can be updated at a normal rate; the closer the value is to 0, the more volatile the system is, and the more necessary it is to slow down or suppress parameter updates.

[0072] For example, within a 5-minute time window, the phase changes of the short-term fluctuation signal are 0.06, 0.05, 0.04, 0.07, and 0.06 radians, respectively, corresponding to an average rate of change of 0.056. During this time period, the phase change direction remains consistent for 80% of the time. Simultaneously, the average energy difference between the short-term signal and the long-term trend signal is 0.12. Based on the established weighting rules, the phase change rate item is weighted at 0.5, the consistency ratio item at 0.3, and the energy difference item at 0.2, resulting in a comprehensive weighted stability coefficient.

[0073] Weight stability coefficient = 0.5×(1−0.056)+0.3×0.8+0.2×(1−0.12) = 0.5×0.944 +0.24+0.176=0.888.

[0074] This result indicates that the model is running relatively stably at this point, and parameter updates can maintain a normal step size. If, within the next time window, the phase change rate suddenly increases to 0.18, the directional consistency drops to 40%, and the energy dissimilarity increases to 0.35, the weight stability coefficient drops to approximately 0.55, meaning that the phase coupling relationship within the model is becoming unstable. At this point, gradient updates should automatically slow down to prevent weight oscillations or parameter collapse. Through this continuous calculation method of the weight stability coefficient, the model can dynamically perceive phase drift and energy fluctuations at different stages of operation, achieving self-adjustment of its internal learning rhythm, thereby maintaining long-term stable predictive performance.

[0075] Through the execution of the above steps, the phase relationship between short-term fluctuation signals and long-term trend signals is transformed from an original static correspondence into a dynamically trackable and sustainably constrained time-continuous mapping structure. The entire process realizes a progressive logic from phase difference extraction to drift gradient analysis, and then to weight stability generation, enabling precise characterization of the temporal relationship between cross-scale signals. Using this implementation method, the model can dynamically adjust the internal parameter update rhythm based on the weight stability coefficient when facing non-stationary grid load changes, effectively avoiding gradient divergence or convergence imbalance.

[0076] Step 3: Based on the weight stability coefficient, an adaptive weight migration mechanism is constructed to perform energy migration and redistribution operations on the weight channels in the gradient phase tracking network that have drift anomalies, forming a multi-scale weight balance structure to maintain the consistency and dynamic balance of weight energy during parameter update.

[0077] Energy transfer and redistribution operations are performed on weight channels exhibiting drift anomalies in the gradient phase tracking network to form a multi-scale weight balancing structure. The specific implementation steps are as follows:

[0078] Based on the aforementioned weight stability coefficients, the energy state of each weight channel in the gradient phase tracking network is continuously monitored and its features are identified. Specifically, each weight channel in the gradient phase tracking network is considered an energy transfer path, and its energy state is characterized by the intensity and direction of parameter changes within a time series. During continuous monitoring, the weight stability coefficients are used as an energy assessment benchmark. By comparing the rate of change and relative magnitude of each weight channel at different times within the time dimension, channels with abnormal energy distribution are identified. When the parameter update intensity of a channel is significantly higher than the average level of adjacent channels, and the duration exceeds a set threshold, it can be determined that the channel exhibits an abnormal drift phenomenon, i.e., excessive energy concentration or directional shift. Conversely, when the update rate of a channel is significantly lower than the overall level and continues to lag, it indicates insufficient energy or delayed transfer. By dynamically comparing all channels, the overall energy distribution map of the network can be obtained, clarifying which channels are in an overloaded state, which channels are in an energy decay state, and which channels are in the equilibrium range. The purpose of this process is to form a continuously updated energy monitoring reference system, providing a basis for judgment in subsequent migration and redistribution operations.

[0079] After identifying weighted channels with abnormal energy distribution, an energy migration operation is performed to establish dynamic energy transfer paths between different channels. Specifically, when an energy overload is detected in a weighted channel, a neighboring channel with high temporal correlation and strong phase stability is preferentially selected as the energy receiving direction. By smoothly adjusting the rate of change of parameters over time, some energy is migrated from the overloaded channel to the under-energy channel. This migration process follows the principle of temporal continuity, ensuring that the energy flow does not disrupt the original phase correspondence, while maintaining the consistency between short-term fluctuation signals and long-term trend signals in the mapping matrix. During the migration process, the weight stability coefficient is still used as an adjustment basis to limit the energy transfer rate and amplitude. When the weight stability coefficient is high, it indicates that the overall phase coupling state is relatively stable. In this case, energy migration should be mainly small adjustments to avoid overcorrection. When the weight stability coefficient is low, it indicates that the system is in a stage of severe phase drift, and the energy migration ratio needs to be increased to restore balance as soon as possible. At the same time, for channels with excessively low energy, if they cannot obtain effective compensation from other channels within a continuous time interval, energy is injected through time series feedback information to gradually restore them to a balanced state with the overall network. This process enables the dynamic transmission of energy in both the time and channel dimensions, so that the weight change process is no longer concentrated in a few highly sensitive channels, but presents an overall coordinated evolution trend, thereby reducing the risk of weight collapse caused by gradient concentration.

[0080] The method for determining neighboring channels with high time correlation and strong phase stability is as follows:

[0081] In a multi-channel weighted structure, several channels adjacent to the target channel are selected as a candidate set. A sliding window correlation analysis is performed on the time series outputs of each candidate channel and the target channel, calculating the Pearson time correlation coefficient and phase synchronization index within that time window. When the correlation coefficient consistently exceeds a preset threshold (e.g., 0.85) and the standard deviation of the phase difference within the entire time window is below a set range (e.g., 0.1 radians), the channel is considered to maintain a high degree of consistency with the target channel in temporal evolution, exhibiting strong temporal correlation and phase stability. Such channels indicate that their internal energy change rhythm is close to that of the target channel, maintaining phase continuity and directional consistency during energy migration. Therefore, they can be preferentially selected as energy receiving channels to ensure smooth energy conduction and dynamic balance.

[0082] After energy transfer, an energy redistribution operation is performed on the overall weighted channels to establish a multi-scale weight balance structure. The key to this step is ensuring that signal characteristics at different time scales are expressed in a balanced manner in the weight space and that a stable energy ratio is maintained during temporal evolution. Specifically, the weighted channels after energy transfer are categorized according to their corresponding time scale characteristics, distinguishing between high-frequency channels that primarily respond to short-term fluctuations and low-frequency channels that primarily represent long-term trends. At the same time baseline, the average energy level of different categories of channels is statistically analyzed to determine if there is a shift in energy distribution. When the energy proportion of high-frequency channels is too high, it indicates that the system is overly sensitive to short-term fluctuations. In this case, the overall energy balance is restored by gradually reducing the energy contribution of high-frequency channels and increasing the weight response of low-frequency channels. When the energy proportion of low-frequency channels is too high, it indicates that the model is too flat in long-term trend fitting, resulting in insufficient response to sudden fluctuations. In this case, the energy weight of high-frequency channels needs to be increased to enhance system sensitivity. Through continuous iteration of this redistribution process, a dynamically balanced energy distribution structure is formed between the high- and low-frequency weighted channels. Throughout the redistribution process, the weight stability coefficient remains as a constraint parameter to ensure that the adjustment process is smooth in time and limited in amplitude, thereby preventing energy from oscillating or becoming disordered again during the redistribution phase. The resulting multi-scale weight equilibrium structure not only achieves the synchronous expression of short-term fluctuation signals and long-term trend signals in the weight space, but also maintains consistency and continuity in the energy flow direction of each channel, providing a stable foundation for subsequent parameter updates.

[0083] Through the above steps, a bidirectional self-balancing mechanism for weights in both the time and channel dimensions is achieved. First, the introduction of a weight stability coefficient enables dynamic sensing capabilities in energy monitoring; second, the energy migration process establishes adaptive energy transfer paths between multiple channels; and finally, energy redistribution ensures the balanced representation of multi-timescale signals within the overall weight structure. This entire process allows the model to proactively adjust its internal weight energy distribution when facing rapid changes in grid load or the superposition of multiple timescale features, preventing excessive gradient concentration in a single channel that could lead to numerical collapse or update instability.

[0084] Step 4: Based on the output of the multi-scale weight balance structure, establish a causal residual playback chain, perform time-series reverse mapping and paired write-back on the residuals output by the power grid load prediction model, and dynamically close-loop correct the parameters of the gradient phase tracking network according to the correlation between historical drift and real-time deviation, thereby realizing self-feedback correction of prediction error during training.

[0085] A causal residual playback chain is established based on the output of a multi-scale weighted equilibrium structure. The specific implementation steps are as follows:

[0086] Based on the output of the multi-scale weighted equilibrium structure, the residuals between the model's predicted values ​​and the actual load observations are obtained, and the temporal characteristics of the residuals are extracted and classified. In the specific implementation, the predicted output within the same time series is compared one by one with the actual grid load at the corresponding time point, and the deviation difference at each time point is calculated, thus forming a continuous residual time series. To ensure that the residuals can truly reflect the prediction error characteristics at different time scales, the residual series is decomposed into short-term fluctuation components and long-term trend components. The short-term component reflects the model's response deviation during rapid fluctuations, while the long-term component reflects the model's systematic error in overall trend prediction. After extraction, the residuals at different time scales are normalized in amplitude and aligned in time to maintain a consistent time coordinate system with the previous cross-scale coupling mapping matrix. Through this process, the residual series not only possesses cross-scale correspondence but also forms a continuous structure in the time dimension that is completely paired with the predicted output, laying the foundation for time matching for subsequent causal playback.

[0087] After obtaining the residual time series, a time-series reverse mapping path for the residuals is established, tracing the prediction error back along the time axis to its generation stage to clarify the causal source of the residuals. Specifically, starting from the residual value at the current moment, the formation process is traced backward along the time axis. By identifying residual abrupt change points, inflection points, and sustained offset intervals, the historical input segments causing the prediction error are determined. During the tracing process, the phase correspondence between short-term fluctuation signals and long-term trend signals in the cross-scale coupling mapping matrix is ​​referenced to analyze the synchronicity and lag between the residuals and signals at different time scales. If the phase change of the residuals and short-term fluctuation signals in a certain time period is synchronized, it indicates that the error mainly stems from insufficient weight adjustment in the high-frequency fluctuation response of the model; if there is a significant lag between the residuals and the changes in the long-term trend signal, it indicates that the error originates from insufficient stability of the model during the long-term trend fitting process. Through this reverse mapping mechanism, the residuals are repositioned to their causal source point in the time dimension, enabling the model to distinguish the formation paths of different types of errors, thereby providing targeted time references and energy direction information for subsequent write-back corrections.

[0088] After identifying the causal sources of the residuals, a paired write-back process is performed, applying them to the parameter space of the multi-scale weighted balancing structure and the gradient phase tracking network. Specifically, the residuals at each time step are assigned to the corresponding time interval and weight channel according to their causal mapping results, ensuring that the write-back direction of the error remains consistent with the original parameter update direction. For high-frequency residuals caused by short-term fluctuations, they are applied to the high-frequency weight channel to correct the sensitivity bias of the model during the fast response phase; for low-frequency residuals caused by long-term trend signals, they are applied to the low-frequency weight channel to correct the smoothness bias of the model during the slow change phase. During the write-back process, the weight stability coefficient acts as a dynamic adjustment factor, constraining the magnitude and rate of residual write-back. A high weight stability coefficient indicates that the current model state is stable, and residual write-back is mainly for fine adjustments; a low weight stability coefficient indicates that the system has significant drift, requiring stronger write-back to achieve rapid correction. By combining causal allocation with dynamic constraints, the residuals are no longer merely errors as statistical values, but are transformed into time feedback signals that drive parameter self-adjustment. This enables the model to proactively perceive and correct its prediction biases at different time scales, thereby achieving a closed-loop flow of error correction.

[0089] After the residuals are paired and written back, the entire causal residual playback chain undergoes temporal continuity maintenance and dynamic closed-loop correction. The goal of this process is to combine residual feedback and parameter update into a closed-loop system with temporal self-stabilizing characteristics. Specifically, the impact of write-back correction on prediction performance is evaluated by periodically monitoring changes in model output before and after residual write-back. If the residual amplitude of the model output decreases significantly and the trend stabilizes over a continuous time period, the closed-loop correction process is considered effective. If the residuals exhibit a new shift trend, the residual propagation path needs to be recalculated based on the latest drift direction, triggering a new round of playback iterations to ensure the feedback process is continuous and adaptive. During this process, the parameters of the gradient phase tracking network are finely adjusted according to the residual write-back signal, allowing the model to gradually converge to the optimal stable range over time. Simultaneously, the causal residual playback chain records the mapping relationship between historical residuals and write-back correction results, forming a temporal memory for subsequent iterative learning. This allows the model to accumulate correction experience, enabling it to adjust in advance when encountering similar load fluctuation patterns in the future. Ultimately, the causal residual replay chain forms a stable feedback loop in time, enabling a continuously self-adjusting dynamic correlation structure between model output, error detection, residual write-back, and parameter correction, thus achieving real-time closed-loop correction of prediction errors in the time dimension.

[0090] This step introduces time-causal logic and a dynamic replay mechanism, enabling the model to have self-stabilization and self-adaptation capabilities. Under complex power grid load change scenarios, it can maintain prediction accuracy and response stability for a long time, providing technical support for achieving highly reliable operation of power grid load prediction and intelligent scheduling.

[0091] Step 5: Based on the stable operation of the causal residual playback chain, a time-scale self-folding mapping is introduced to project short-term fluctuation signals and long-term trend signals to different time and frequency domains, respectively, to construct a fast-response prediction kernel and a slow-change trend kernel. The phase conjugate operator is used to perform energy synchronization correction on the prediction kernel, so that the power grid load prediction model automatically splits into independent sub-models when phase resonance is detected, so as to prevent gradient interference between signals of different time scales from the source and maintain the dynamic stability of the prediction model.

[0092] Based on the stable operation of the causal residual replay chain, a time-scale self-folding mapping is introduced. The specific implementation steps are as follows:

[0093] First, based on the stable operation of the causal residual playback chain, time-scale hierarchical and self-folding mapping is performed on the time output sequence of the multi-scale weighted equilibrium structure. Specifically, the model output sequence after residual feedback correction is obtained first, based on the time closed-loop correction results formed in the causal residual playback chain. This output sequence has achieved phase coordination and energy balance of multi-scale signals in the time dimension, and therefore can be used as the initial input for time-folding mapping. Subsequently, based on the rate of time change and the frequency of energy change, the entire time series is divided into short-time-scale segments and long-time-scale segments. The short-time-scale segments mainly reflect the rapid disturbance characteristics of the power grid load, such as sudden weather changes, instantaneous fluctuations in new energy sources, or short-term surges in user electricity consumption; the long-time-scale segments describe the slow evolution trend of the overall load, such as diurnal cycles, weekday patterns, and seasonal fluctuations. By establishing a time-scale mapping relationship, the original time axis is folded into two interconnected sub-time domains, allowing short-time and long-time signals to unfold independently within their respective time-folding spaces, while maintaining phase continuity and energy conservation at the mapping boundaries. This process forms a dynamic mapping structure in two time domains, providing a continuous time basis for subsequent frequency domain projection and energy decomposition.

[0094] After completing the time-scale self-folding mapping, short-term fluctuation signals and long-term trend signals are projected onto different time-frequency domains to achieve the separation and independent response of characteristic energies. Specifically, along the time axis formed by the self-folding mapping, high-frequency features of the short-term fluctuation signal are identified and projected into the high-frequency domain, preserving primarily transient changes and short-term fluctuation characteristics. Simultaneously, the long-term trend signal is projected into the low-frequency domain according to its slow-changing characteristics, representing the stable change pattern of the long-term trend. During the projection process, a continuous transition between the two types of signals at the time boundaries is maintained, ensuring a natural connection in the energy transfer between the short-term and long-term signals. After this processing, the short-term fluctuation signal and the long-term trend signal evolve independently in the high-frequency and low-frequency domains, respectively, forming relatively stable time-frequency distribution structures, thus avoiding direct interference between signals of different time scales in the same time domain. This step achieves the physical separation of signal energy in the time-frequency dimension, laying the foundation for constructing prediction kernels with different response characteristics.

[0095] After separating the signals in the high and low frequency domains, a fast-response prediction kernel and a slow-change trend kernel are constructed to characterize the independent features of load evolution at different time scales. Specifically, in the high-frequency domain, a fast-response prediction kernel with high sensitivity and fast response characteristics is constructed based on the energy distribution and phase change rate of short-time fluctuation signals, enabling it to capture high-dynamic events such as instantaneous load changes, wind and solar power output fluctuations, or local disturbances. In the low-frequency domain, a slow-change trend kernel with stable output and slow evolution characteristics is constructed based on the stable change law of long-term trend signals, enabling it to continuously describe the long-term trend and structural evolution of the overall power grid load. To maintain energy coordination between the two prediction kernels, the boundary points of the time folding mapping are used as energy transfer nodes during the construction process, enabling continuous energy transition and information exchange between the high-frequency and low-frequency prediction kernels at the boundary, thereby ensuring the overall energy conservation and structural integrity of the model. Through this process, short-time and long-term signals are not only effectively separated, but corresponding response expressions are also obtained in different frequency domains, enabling the model to handle both fast and slow-change characteristics simultaneously.

[0096] Based on the establishment of fast-response prediction kernels and slow-change-trend kernels, a phase conjugation operator is introduced to synchronously correct the energy distribution and phase relationship between the two types of prediction kernels. Specifically, the phase change difference between high-frequency and low-frequency prediction kernels in the boundary region is first detected, and the phase shift direction and energy difference amplitude at the time-folding mapping boundary are calculated. If the high-frequency and low-frequency prediction kernels are found to amplify or cancel each other's energy in this region, it indicates a phase resonance trend. In this case, the phase conjugation operator is used to conjugate and reverse the phases of the two types of prediction kernels at the boundary, ensuring that the phase of the high-frequency prediction kernel and the low-frequency prediction kernel maintain opposite energy but complementary directions, thereby eliminating the energy resonance phenomenon caused by phase superposition. Simultaneously, the energy distribution within the prediction kernels is synchronously corrected, ensuring that the energy change rates of the high-frequency and low-frequency domains remain balanced within the continuous interval of the time-folding mapping, preventing energy concentration on a single time scale and causing instability in the model output. Through this step, short-time and long-term signals achieve complementary coordination at the phase and energy levels, laying the physical foundation for preventing gradient interference.

[0097] Finally, after phase conjugate correction, when the model detects significant phase resonance or enhanced energy coupling between high and low frequency signals in the time-frequency domain, it triggers an automatic splitting operation, decomposing the overall model into multiple independent sub-models. Specifically, when short-term fluctuation signals and long-term trend signals exhibit sustained phase synchronization and energy amplification effects within a certain region of the time-folding map, to prevent nonlinear accumulation of gradients in this region, the high-frequency sub-model and low-frequency sub-model in this region are separated through a splitting mechanism, allowing each to operate in an independent time-frequency sub-domain. The split sub-models inherit the phase conjugate relationship and energy correction mechanism of the original model, enabling them to evolve independently on their respective time scales and maintain local stability. As time progresses, when the system re-enters a stable phase and the phase difference returns to the normal range, the sub-models can be recombined into a complete structure, thus achieving a dynamically separable and combinable adaptive modeling process. Through this bidirectional adjustment mechanism of splitting and merging, the model can autonomously identify phase resonance and perform timely structural separation when facing complex multi-scale power grid load signals. This eliminates gradient interference and energy disturbance between signals of different time scales from the source, ensuring that the entire prediction process maintains a stable and controllable evolutionary state in the three-dimensional space of time, phase, and energy.

[0098] This process starts with the continuity of the time scale, focuses on frequency domain energy coordination, and ends with adaptive decomposition of the model structure. It enables short-term fluctuation signals and long-term trend signals to achieve domain independence and energy balance within a unified spatiotemporal framework. Through this mechanism, the model can proactively identify potential phase resonances and self-reconstruct in highly dynamic power grid operating environments, thereby maintaining the continuity, stability, and physical consistency of the prediction output over the long term. This provides a solution for realizing smart grid load forecasting and adaptive scheduling.

[0099] This invention introduces a multi-scale dynamic baseline and gradient phase tracking mechanism into the load forecasting process, enabling the model to achieve continuous alignment in the time dimension and adaptive phase difference correction when facing the interaction of short-term fluctuation signals and long-term trend signals. This process establishes a stable phase mapping relationship between short-term and long-term features under a unified time coordinate. During the model update process, the gradient step size is automatically adjusted according to the weight stability coefficient, avoiding gradient oscillations and parameter collapse caused by multi-timescale drift. By maintaining the dynamic balance of weight energy, the model can maintain structural stability and output consistency during the operational phase where rapid load fluctuations and gradual trend changes occur alternately, thereby significantly improving the temporal continuity and dynamic accuracy of load forecasting.

[0100] This invention constructs a causal residual playback chain and a time-scale self-folding mapping mechanism, enabling the model to possess closed-loop self-correction capabilities for historical biases and real-time errors. Furthermore, it actively triggers sub-model splitting upon detecting phase resonance between multi-scale signals, achieving adaptive steady-state adjustment at the structural level. This mechanism allows the model to identify the sources of prediction biases in real-time and perform energy synchronization correction in complex power grid operating environments, ensuring that prediction responses at different time scales maintain independent and coordinated evolutionary characteristics. Through this time-series closed-loop and structural self-separation approach, the model possesses continuous self-stability and anti-interference capabilities during long-term operation, effectively reducing the risk of scheduling link instability and improving the stability and overall operational efficiency of power grid load forecasting and scheduling.

[0101] This invention provides, for example Figure 4 The illustrated zonal optimization scheduling and control system for an energy storage power station includes a multi-scale dynamic baseline construction module, a gradient phase tracking module, an adaptive weight transfer module, a causal residual playback module, and a time-scale self-folding module.

[0102] The multi-scale dynamic baseline construction module constructs a multi-scale dynamic baseline for load data, performs phase calibration on short-term fluctuation signals and long-term trend signals under a unified time coordinate, and generates a cross-scale coupling mapping matrix based on the calibration results using a time embedding operator.

[0103] The gradient phase tracking module establishes a gradient phase tracking network based on a cross-scale coupling mapping matrix, continuously analyzes the phase difference between short-term fluctuation signals and long-term trend signals, extracts phase drift gradient information, and calculates and generates weight stability coefficients based on the phase drift gradient.

[0104] The adaptive weight transfer module builds an adaptive weight transfer mechanism based on the weight stability coefficient. It performs energy transfer and redistribution operations on weight channels that have drift anomalies in the gradient phase tracking network, forming a multi-scale weight balance structure.

[0105] The causal residual playback module establishes a causal residual playback chain based on the output results of the multi-scale weight balance structure. It performs time-series reverse mapping and paired write-back on the residuals output by the power grid load prediction model, and dynamically closes the loop to correct the parameters of the gradient phase tracking network according to the correlation between historical drift and real-time deviation.

[0106] The time-scale self-folding module, based on the stable operation of the causal residual playback chain, introduces time-scale self-folding mapping to project short-term fluctuation signals and long-term trend signals to different time and frequency domains respectively, constructing a fast-response prediction kernel and a slow-change trend kernel, and using the phase conjugate operator to perform energy synchronization correction on the prediction kernel, so that the power grid load prediction model automatically splits into independent sub-models when phase resonance is detected.

[0107] The present invention provides a partitioned optimal scheduling and control method for an energy storage power station, which is implemented through the partitioned optimal scheduling and control system for an energy storage power station described above. For details of the specific method and process of the partitioned optimal scheduling and control system for an energy storage power station, please refer to the embodiment of the partitioned optimal scheduling and control method for an energy storage power station described above, which will not be repeated here.

[0108] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for partition optimization scheduling control of an energy storage power station, characterized in that, The method comprises the following steps: Step one, constructing a multi-scale dynamic baseline of load data, phase calibration of short-term fluctuation signals and long-term trend signals in a unified time coordinate, and generating a cross-scale coupling mapping matrix based on the calibration results using a time embedding operator; Step two, establishing a gradient phase tracking network based on the cross-scale coupling mapping matrix, continuously analyzing the phase difference between the short-term fluctuation signals and the long-term trend signals, extracting the phase shift gradient information, and calculating the weight stability coefficient according to the phase shift gradient; Step three, constructing an adaptive weight migration mechanism based on the weight stability coefficient, performing energy migration and redistribution operations on the weight channels that appear to be abnormal in the gradient phase tracking network, and forming a multi-scale weight balance structure; Step four, based on the output results of the multi-scale weight balance structure, establishing a causal residual playback chain, performing time sequence reverse mapping and paired write-back on the residual error output by the power grid load prediction model, and dynamically closing the loop to modify the parameters of the gradient phase tracking network according to the correlation between historical drift and real-time deviation; Step five, on the basis of stable operation of the causal residual playback chain, introducing time scale self-folding mapping, projecting the short-term fluctuation signals and the long-term trend signals into different time frequency domains respectively, constructing fast response prediction kernel and slow change trend kernel, and using phase conjugate operator to perform energy synchronous correction on the prediction kernel, so that the power grid load prediction model automatically splits into independent sub-models when phase resonance is detected.

2. The partition optimization scheduling control method of the energy storage power station according to claim 1, characterized in that, The step of constructing a multi-scale dynamic baseline of load data specifically includes: Multi-scale decomposition and time unified mapping of original power grid load data, extracting key elements from historical power grid operation data, and dividing the complete load sequence into short-term fluctuation signal sequence and long-term trend signal sequence through unified time resolution sampling and time interval layering, and performing time scale correction on the short-term fluctuation signal and the long-term trend signal; After establishing a unified time coordinate, phase calibration of short-term fluctuation signals and long-term trend signals, determination of inflection point position by calculating the change direction and rate of short-term fluctuation signals and long-term trend signals, and smoothing of phase difference in a sliding time window to form a continuous time mapping sequence; After obtaining the phase calibration results, a cross-scale coupling mapping matrix is generated using a time embedding operator, the phase correspondence is taken as the time embedding input, a set of time nodes is set on the unified time axis, and a continuous interpolation relationship is established, so that different time scale signals remain consistent in phase, amplitude and energy distribution.

3. The partition optimization scheduling control method of an energy storage power station according to claim 2, characterized in that, The step of calculating the weight stability coefficient according to the phase shift gradient includes: Time domain expansion and phase difference extraction of the generated cross-scale coupling mapping matrix, extraction of phase correspondence information of short-term fluctuation signals and long-term trend signals at each time, identification of phase change direction and rate between adjacent time points, and formation of a phase shift curve of short-term fluctuation signals relative to long-term trend signals; After obtaining the continuous phase difference data, the variation characteristics of the phase shift curve are analyzed and identified to determine the duration and amplitude of the rising, falling and stable intervals, and a phase shift time correlation sequence reflecting the interaction characteristics of short-term fluctuation signals and long-term trend signals is formed; Based on the phase shift time correlation sequence, phase shift gradient information is extracted, and a gradient change reference line is established according to the time continuous change trend to determine the phase shift direction and persistence, and a dynamic gradient input representing the phase coupling stability is formed; According to the extracted phase shift gradient information, a weight stability coefficient is calculated and generated, and the phase stability degree is determined by analyzing the amplitude change and direction switching frequency of the phase shift gradient. When the shift gradient is gentle and the direction is consistent, a high weight stability coefficient is generated. When the shift gradient fluctuates sharply, a low weight stability coefficient is generated.

4. The partition optimization scheduling control method of an energy storage power station according to claim 3, characterized in that, In the step of extracting the phase shift gradient information, the direction consistency and change rate of the phase shift time correlation sequence in the continuous time interval are comprehensively judged. Only when the phase shift direction remains continuous within the preset time threshold and the change amplitude is stable, the shift amount of the continuous time interval is counted into the effective gradient range.

5. The partition optimization scheduling control method of an energy storage power station according to claim 3, characterized in that, The steps of performing energy migration and redistribution operation on the weight channel with drift anomaly in the gradient phase tracking network to form a multi-scale weight balance structure include: Based on the weight stability coefficient, the energy state of each weight channel in the gradient phase tracking network is continuously monitored and characterized. By comparing the channel parameter change rate and relative amplitude at different times, the channel with abnormal energy distribution is identified and the overall energy distribution map is formed; After identifying the weight channel with abnormal energy distribution, an energy migration operation is performed. The adjacent channel with high time correlation and strong phase stability is selected as the energy receiving direction. The energy is dynamically conducted between the channels by smoothing the parameter change rate, and the weight stability coefficient is used to limit the migration rate and amplitude to maintain time continuity; After completing the energy migration, an energy redistribution operation is performed on each weight channel. According to the time scale characteristics of the channel, high-frequency channels and low-frequency channels are distinguished. According to the energy distribution state, the channel energy proportion is adjusted to realize the synchronous expression of short-term fluctuation signals and long-term trend signals in the weight space and maintain the dynamic balance of energy.

6. The partition optimization scheduling control method of an energy storage power station according to claim 5, characterized in that, In the energy redistribution operation, the weight stability coefficient is used to dynamically constrain the energy adjustment amplitude of high-frequency channels and low-frequency channels. When the weight stability coefficient is higher than the set threshold, small amplitude energy adjustment is performed to maintain channel balance. When the weight stability coefficient is lower than the set threshold, the energy migration proportion is increased to accelerate the recovery of balance.

7. The partition optimization scheduling control method of an energy storage power station according to claim 5, characterized in that, The steps of establishing a causal residual playback chain based on the output results of the multi-scale weight balance structure include: Based on the output results of the multi-scale weight balance structure, the residual between the model prediction value and the actual load observation value is obtained, and the residual is decomposed and time-aligned into short-term fluctuation components and long-term trend components to form a residual time series with cross-scale correspondence; After obtaining the residual time series, a time reverse mapping path of the residual is established to track the prediction error in reverse along the time axis to its formation stage, and the causal source of the residual is determined according to the phase correspondence between the short-term fluctuation signal and the long-term trend signal; After determining the causal source of the residual, the residual is subjected to paired write-back processing, which is applied to the corresponding high-frequency weight channel or low-frequency weight channel, and the amplitude and rate of residual write-back are dynamically constrained according to the weight stability coefficient; After completing the paired write-back of the residual, the time continuity of the causal residual playback chain is maintained and closed-loop correction is performed by periodically monitoring the residual change trend and the write-back effect to form a time closed-loop structure among model output, error detection and parameter correction.

8. The partition optimization scheduling control method of an energy storage power station according to claim 7, characterized in that, The steps of introducing time scale self-folding mapping on the basis of stable operation of the causal residual playback chain include: On the basis of stable operation of the causal residual playback chain, the time output sequence of the multi-scale weight balance structure is subjected to time scale layering and self-folding mapping, and the original time axis is folded into a short time scale section and a long-term time scale section; The short-term fluctuation signal and the long-term trend signal after self-folding mapping are projected into different time frequency domains, so that the short-term fluctuation signal is projected into the high frequency domain and the long-term trend signal is projected into the low frequency domain; After high and low frequency domain separation, fast response prediction kernel and slow change trend kernel are respectively constructed, and energy transfer nodes are established at the boundary points of time folding mapping; The phase conjugate operator is introduced to simultaneously correct the energy distribution and phase relationship of the high and low frequency prediction kernels, so that they are phase complementary at the boundary to eliminate the phase resonance phenomenon; When it is detected that there is persistent phase resonance or energy coupling enhancement between high and low frequency signals in the time frequency domain, the power grid load prediction model structure is triggered to automatically split, and the overall model is decomposed into independent sub-models.

9. A subarea optimization scheduling control system of an energy storage power station, used for implementing the subarea optimization scheduling control method of the energy storage power station according to any one of claims 1-8, characterized in that, It includes a multi-scale dynamic baseline construction module, a gradient phase tracking module, an adaptive weight migration module, a causal residual playback module, and a time scale self-folding module; The multi-scale dynamic baseline construction module constructs a multi-scale dynamic baseline of the load data, phase labels the short-term fluctuation signal and the long-term trend signal in a unified time coordinate, and generates a cross-scale coupling mapping matrix based on the labeling results using a time embedding operator; The gradient phase tracking module establishes a gradient phase tracking network based on the cross-scale coupling mapping matrix, continuously analyzes the phase difference between the short-term fluctuation signal and the long-term trend signal, extracts the phase drift gradient information, and calculates the weight stability coefficient according to the phase drift gradient; The adaptive weight migration module relies on the weight stability coefficient to construct an adaptive weight migration mechanism, performs energy migration and redistribution operations on the weight channels that appear drift anomalies in the gradient phase tracking network, and forms a multi-scale weight balance structure; The causal residual playback module establishes a causal residual playback chain based on the output results of the multi-scale weight balance structure, performs time reverse mapping and paired write-back on the residual of the power grid load prediction model output, and dynamically closes the loop to correct the parameters of the gradient phase tracking network according to the correlation between historical drift and real-time deviation. The time scale self-folding module introduces time scale self-folding mapping on the basis of stable operation of the causal residual error playback chain, projects short-time fluctuation signals and long-term trend signals into different time frequency domains respectively, constructs fast response prediction kernels and slow change trend kernels, and performs energy synchronous correction on the prediction kernels by using a phase conjugate operator, so that the power grid load prediction model is automatically split into independent sub-models when phase resonance is detected.

Images