Energy consumption prediction and scheduling control method based on machine learning

Abstract

The invention discloses an energy consumption prediction and scheduling control method based on machine learning. According to the method, operation parameters, energy consumption curves and environment disturbance data of multiple devices are collected through a distributed sensing terminal, the data are input into a pre-trained machine learning model, and a probability prediction result of future energy consumption distribution is generated. On the basis of prediction, an intervention signal is applied in an equipment safety boundary, equipment response characteristics are obtained according to the difference before and after intervention, and an energy consumption risk map is constructed by combining the equipment response characteristics with a probability prediction result. And based on the energy consumption risk map, generating an extreme disturbance scene by using digital twinning, performing consistency check on a probability prediction result and a scheduling scheme in a data domain and a physical domain, and performing multi-stage scheduling in combination with task delays to generate a scheduling result. And finally, issuing the scheduling result to the equipment. The method can improve the accuracy of energy consumption prediction and the reliability of scheduling decision making, and is suitable for intelligent management of data centers, industrial production and high-energy-consumption scenes.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a machine learning-based method for energy consumption prediction and scheduling control. Background Technology

[0002] In existing technologies, energy consumption prediction and scheduling control methods are mostly based on statistical modeling or machine learning modeling using historical energy consumption curves and real-time monitoring data, combined with certain scheduling rules or optimization algorithms to achieve energy resource allocation and control. In practical applications, commonly used techniques include energy consumption prediction based on time-series analysis, load trend judgment using deep learning models, and energy consumption optimization through demand response strategies and hierarchical scheduling mechanisms. These methods have been widely applied in smart grids, data center management, and industrial production, providing fundamental support for improving energy utilization efficiency.

[0003] However, existing technologies still have many shortcomings when dealing with energy consumption prediction and scheduling in complex environments. First, prediction models mostly rely on static historical data, lacking sensitivity to sudden disturbances and dynamic changes in equipment operating status, making it difficult to detect potential risks in advance. Second, energy consumption scheduling is mostly based on a single optimization objective, lacking a comprehensive consideration of equipment health status and task flexibility, resulting in limitations in the reliability and adaptability of scheduling results. In addition, existing methods lack a mechanism for multi-domain consistency verification of prediction and scheduling results, leading to discrepancies between prediction and execution, making it difficult to achieve closed-loop optimization.

[0004] Therefore, a new method for energy consumption prediction and scheduling control that can integrate prediction accuracy, risk assessment and scheduling optimization is needed. Summary of the Invention

[0005] This application provides a machine learning-based energy consumption prediction and scheduling control method to improve the accuracy of energy consumption prediction and the reliability of scheduling decisions.

[0006] This application provides a machine learning-based energy consumption prediction and scheduling control method, including: The system collects operating parameters, energy consumption curves, and environmental disturbance data from multiple devices through distributed sensing terminals. The operating parameters, energy consumption curves, and environmental disturbance data are input into a pre-trained machine learning model, which combines a temporal recursive structure and attention weights to generate a probability prediction result of energy consumption distribution. An intervention signal is applied within the equipment safety boundary, and equipment response characteristics are generated based on the intervention differences. These characteristics are then combined with probability prediction results to form an energy consumption risk map. Based on the energy consumption risk map, extreme disturbance scenarios are generated using digital twins. The consistency between the probability prediction results and the scheduling scheme is verified in the data domain and physical domain. Multi-level scheduling is performed based on the verified energy consumption risk map and the task delayability to form the scheduling result. The scheduling results are sent to the devices and feedback data is transmitted back. The machine learning model is then updated based on the feedback data.

[0007] The beneficial effects of the technical solution provided in this application include: (1) By combining a machine learning model with a temporal recursive structure and attention weights, the prediction process can take into account both long-term energy consumption trends and short-term fluctuation characteristics, thereby improving the accuracy and robustness of energy consumption distribution prediction. (2) By applying intervention signals within the equipment safety boundary and generating equipment response characteristics, the energy consumption risk map can reflect the actual dynamic behavior of the equipment in response to disturbances, thereby identifying potential energy consumption anomalies and operational risks in advance. (3) By using digital twins to generate extreme disturbance scenarios and verifying the consistency of prediction results and scheduling schemes in the data domain and physical domain, the scheduling scheme can be ensured to conform to both data statistical laws and physical feasibility, thereby improving the reliability of scheduling decisions. (4) By introducing multi-level scheduling through task delayability and updating the machine learning model with feedback data, a closed-loop control of prediction, evaluation and execution can be formed, thereby achieving continuous optimization and adaptive evolution of energy consumption management. Attached Figure Description

[0008] Figure 1 This is a flowchart of a machine learning-based energy consumption prediction and scheduling control method provided in the first embodiment of this application. Detailed Implementation

[0009] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0010] The first embodiment of this application provides a machine learning-based energy consumption prediction and scheduling control method. Please refer to... Figure 1 This figure is a schematic diagram of the first embodiment of this application. The following is in conjunction with... Figure 1 The first embodiment of this application provides a detailed description of an energy consumption prediction and scheduling control method based on machine learning.

[0011] Step S101: Collect operating parameters, energy consumption curves and environmental disturbance data of multiple devices through distributed sensing terminals.

[0012] In step S101, distributed sensing terminals need to be deployed in the target application scenario. Each sensing terminal should have the ability to continuously monitor and collect data from a specified object. The sensing terminals may include voltage sensors, current sensors, power factor sensors, temperature sensors, humidity sensors, and environmental noise or vibration sensors, etc., to comprehensively acquire the status parameters of the device during operation. The operating parameters should at least include the device's voltage value, current value, active power, reactive power, power factor, and frequency information, so that the subsequent model can accurately characterize the device's load characteristics and energy usage. The energy consumption curve is time-series data obtained through sampling and cumulative calculation based on the above electrical parameters. For example, load curves and electricity consumption curves sampled at the second or minute level can be stored as a serialized data structure with timestamps to ensure data continuity and traceability.

[0013] In addition to equipment operating parameters and energy consumption curves, it is also necessary to collect external environmental disturbance data. This type of data describes external factors that may affect energy consumption fluctuations. Environmental disturbance data includes, but is not limited to, temperature, humidity, atmospheric pressure, airflow speed, ambient light intensity, and start-up or shutdown events at the production site. This information is acquired through environmental sensors, PLC signal acquisition modules, or external data interfaces. For example, in industrial production scenarios, process start-up and shutdown signals and work order execution status can be directly collected by interfacing with the production control system and stored in correspondence with real-time energy consumption curves.

[0014] In practical implementation, each distributed sensing terminal needs to convert the original analog signal into a digital signal through a local sampling module. The sampling precision can be set to 16 bits or higher, and the sampling frequency can be set from 1kHz to 10kHz depending on the device characteristics to ensure the ability to capture instantaneous fluctuations. The acquired digital signal needs to undergo local preprocessing, including filtering and noise reduction, baseline correction, time synchronization, and data compression, to reduce transmission bandwidth pressure. Time synchronization can be achieved through GPS time synchronization or network time protocols to ensure that the data collected by all distributed sensing terminals are aligned at the millisecond level, avoiding timing errors from causing deviations in subsequent model training and prediction.

[0015] During data transmission, each distributed sensing terminal uploads pre-processed data to the data aggregation node or central server via a wired communication interface (such as Ethernet, RS485) or a wireless communication protocol (such as Wi-Fi, 4G / 5G, LoRa). To ensure that data is not tampered with or lost during transmission, a checksum, such as a CRC checksum or hash digest, can be generated for each set of data at the sensing terminal side and verified at the receiving end. If an inconsistency is found, a retransmission mechanism is triggered. To improve data security, TLS encryption or lightweight encryption algorithms can be superimposed at the transmission protocol layer to ensure the integrity and confidentiality of data during transmission.

[0016] In the data storage stage, the collected operating parameters, energy consumption curves, and environmental disturbance data need to be archived according to a unified data format. A time-series database-based storage structure can be adopted, such as InfluxDB or other databases suitable for high-frequency time-series data, storing all data indexed by device ID, parameter type, and timestamp. This ensures that when calling machine learning models later, the complete data sequence of any device within any time period can be directly extracted, achieving fast data retrieval.

[0017] To facilitate subsequent processing, edge computing functionality can be configured at the data aggregation node to fuse data from different sensing terminals. For example, multi-source data synchronization alignment, outlier removal, and missing value imputation can be performed at the edge to ensure the consistency and high quality of data input into subsequent models. For missing data, linear interpolation, Kalman filtering, or imputation methods based on similar device data can be used for filling; for outliers, statistical thresholds can be set or algorithms such as isolated forests can be used to automatically identify and remove them, thereby ensuring the stability and accuracy of the overall dataset.

[0018] In summary, the implementation process of step S101 covers the entire process of sensor terminal deployment and content definition, data sampling and preprocessing, time synchronization and transmission assurance, data storage and index management, and multi-source data fusion and cleaning.

[0019] Furthermore, the step of collecting operating parameters, energy consumption curves, and environmental disturbance data of multiple devices through distributed sensing terminals includes: In the distributed sensing terminal, the voltage, current and power signals of the device are sampled at high frequency, and an initial sequence with timestamps is generated during the sampling process; In the distributed sensing terminal, multi-scale wavelet decomposition is performed on the initial sequence to extract energy consumption fluctuation components at different time scales, and the energy consumption fluctuation components are reconstructed into an energy consumption multi-resolution feature sequence. In a distributed sensing terminal, the energy consumption multi-resolution feature sequence is coupled with environmental disturbance parameters to construct an environmental labeled energy consumption vector. The environmental labeled energy consumption vector contains operating parameter features and environmental disturbance weights within the same time slice. In the distributed sensing terminal, a dynamic reliability assessment is performed on the environmental labeled energy consumption vector. A reliability factor is calculated by performing a deviation test with the historical distribution, and the reliability factor is appended to the environmental labeled energy consumption vector to form a reliable energy consumption sample. The data aggregation node receives the trusted energy consumption samples and performs time alignment and spatial fusion on the trusted energy consumption samples from different distributed sensing terminals, outputting a high-integrity energy consumption dataset in a unified format for subsequent machine learning models to call.

[0020] In this embodiment, multiple distributed sensing terminals are deployed on or near different devices, each with independent data acquisition and preprocessing capabilities. The device's operating parameters are primarily acquired in real-time via voltage, current, and power sensors. A high-frequency sampling mode is used during acquisition, with the sampling frequency set at the kilohertz level to ensure the capture of subtle fluctuations in the device over a short period. Each sampling result is accompanied by a precise timestamp, and synchronization calibration between the internal and global clocks ensures accurate time alignment of data from all distributed terminals. The resulting initial sequence is a set of multi-dimensional data points arranged continuously over time, each data point including voltage, current, and power values, along with corresponding timestamp information.

[0021] To further extract energy consumption features at different time scales, the sensing terminal performs multi-scale wavelet decomposition on the initial sequence. Specifically, the sensing terminal divides the signal into high-frequency and low-frequency components, corresponding to short-period bursts in energy consumption and long-period stable trends in energy consumption. Through multiple decompositions, multiple levels of energy consumption fluctuation components can be obtained; for example, the first level represents second-level fluctuations, the second level represents minute-level fluctuations, and the third level represents hour-level fluctuations. After recombinating these components, a multi-resolution energy consumption feature sequence is obtained. This sequence is constructed by simultaneously retaining feature values ​​at different scales at each time point, enabling subsequent processing to not only observe the overall trend but also capture instantaneous anomalies.

[0022] After generating the multi-resolution energy consumption feature sequence, it needs to be coupled with environmental disturbance parameters. These parameters include external data such as temperature, humidity, ambient noise, and airflow velocity. The core of the coupling process is to bind energy consumption features to environmental disturbance parameters within the same time slice. For example, at a certain time point, if the multi-resolution energy consumption feature sequence shows high short-term fluctuations, and the ambient temperature is high while humidity is rising, then the environmentally labeled energy consumption vector at that time point includes both operational parameter features and environmental disturbance weights. The environmental disturbance weights are calculated based on the magnitude of the environmental disturbance changes and their correlation with energy consumption; for example, a rapid rise in temperature may be assigned a higher weight, while a stable change in humidity may be assigned a lower weight. This process ensures that each vector reflects not only the device's own operating status but also the impact of external conditions on energy consumption behavior.

[0023] After generating the environmental labeled energy consumption vector, a dynamic reliability assessment needs to be performed. The assessment compares the currently collected environmental labeled energy consumption vector with its historical distribution to determine the magnitude of the deviation. If the current data point differs little from the historical average and the fluctuation range is within the normal range, a higher reliability factor is assigned. Conversely, if a data point experiences an abnormal increase in current or atypical sudden fluctuations in power, which are rare in historical records, a lower reliability factor is assigned. The reliability factor is not calculated based on a single indicator but through a comprehensive evaluation of multiple dimensions, such as mean deviation, variance deviation, and distribution range coverage. The calculation result is appended as a value to the environmental labeled energy consumption vector, thus forming a reliable energy consumption sample. The advantage of the reliable energy consumption sample is its ability to distinguish normal data from potentially noisy or abnormal data.

[0024] Finally, these trusted energy consumption samples are transmitted to the data aggregation node. At the aggregation node, the samples uploaded from different sensing terminals are first time-aligned, and all samples are arranged according to precise timestamps using a unified global time base. Then, spatial fusion is performed, combining trusted energy consumption samples from different devices within the same time slice to form a multi-device joint sample. For example, the voltage, current, and power characteristics of multiple servers in a data center within the same time slice are aggregated into a single data matrix, which also includes temperature and humidity information from the same time. In this way, the highly complete energy consumption dataset output by the data aggregation node not only preserves the independent characteristics of each device but also reflects the collaborative relationships between devices. The dataset is standardized to a standard format, ensuring that subsequent machine learning models can directly use it without additional preprocessing.

[0025] This approach ensures that the entire process, from raw data acquisition to the generation of the final high-integrity energy-efficient dataset, is feasible.

[0026] Step S102: Input the operating parameters, energy consumption curve and environmental disturbance data into the pre-trained machine learning model. The machine learning model combines the temporal recursive structure and attention weights to generate a probability prediction result of energy consumption distribution.

[0027] In step S102, the collected operating parameters, energy consumption curves, and environmental disturbance data need to be organized and preprocessed according to a unified data format to ensure that they can be correctly accepted and processed by the machine learning model. First, the raw data should be normalized or standardized to map data with different dimensions and value ranges to a unified numerical interval. For example, input parameters such as voltage, current, temperature, and humidity can be linearly scaled to the [0,1] or [-1,1] interval to avoid unreasonable weight biases in model calculations caused by parameters with excessively large values. When handling missing values, mean interpolation based on historical sequences, adjacent interpolation, or extrapolation using correlations with other sensing terminals can be used to ensure the integrity of the input sequence in the time dimension. For outliers that significantly exceed the normal range, threshold judgment, statistical distribution testing, or the isolated forest algorithm are needed for identification, and they should be removed or corrected as necessary to ensure the quality of the input data.

[0028] After the data is cleaned and formatted, an input sequence needs to be constructed in chronological order, and each sampling time point is associated with corresponding operating parameters, energy consumption curve data, and environmental disturbance information. This input sequence, as a multi-dimensional time series tensor, is fed into a pre-trained machine learning model. The core structure of the model is a hybrid architecture combining a temporal recurrent network and an attention mechanism. The temporal recurrent network can use a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU) to extract long-term dependencies and short-term fluctuation features from the input sequence. This part uses a gating unit to retain important historical states and suppress irrelevant or noisy information during information transmission, enabling the model to effectively characterize the periodicity and non-stationarity of the equipment's energy consumption curve.

[0029] An attention mechanism is introduced on top of the recursive structure. Specifically, at each time step, the correlation score between the current hidden state and the hidden vectors at all times in the input sequence is calculated. This score is then normalized to obtain the attention weight, which is used to weight and summarize the features at each time step. In this way, the model not only relies on the sequential memory of the recursive network but also automatically focuses on the time segments and feature variables that have the greatest impact on the prediction results through the attention mechanism, thereby strengthening key perturbation signals and ignoring minor fluctuations.

[0030] During the training phase, the hybrid model iteratively optimizes using a large amount of historical energy consumption data. The loss function can employ probability distribution-related metrics, such as negative log-likelihood or quantile loss, enabling the model to learn to output not just single-point predictions, but a complete probability distribution. After training, in the inference phase, real-time collected and preprocessed operating parameters, energy consumption curves, and environmental disturbance data are input into the model. The model first captures the temporal patterns of the data through a recursive structure, then highlights key factors through an attention mechanism, and finally generates energy consumption distribution predictions for future periods at the output layer. These predictions are given in probabilistic form, such as providing mean predictions, variances, and the distribution of possible energy consumption values ​​within a certain confidence interval, or directly outputting multiple quantile predictions to characterize various possible future energy consumption scenarios.

[0031] To make the probabilistic prediction results usable, a softmax layer or a parameterized output method based on Gaussian distribution or quantile regression can be introduced in the output stage. This allows the model's prediction of future energy consumption to be no longer limited to a single value, but to form distribution information with uncertainty description.

[0032] Furthermore, the machine learning model includes a temporal decomposition unit, a two-stream recursive unit, a mask attention unit, a constraint-based fusion unit, and a probability flow generation unit; The time series decomposition unit is used to receive operating parameters, energy consumption curves and environmental disturbance data, perform timestamp alignment and multi-scale segmentation on the received data, and obtain an initial interval feature sequence. The time series decomposition unit is implemented by wavelet decomposition and sliding window method, so that energy consumption fluctuations and long-term trends at different time scales can be preserved in the initial interval feature sequence at the same time. The dual-stream recursive unit is used to receive the initial interval feature sequence, generate steady-state hidden states and perturbation hidden states respectively, and obtain a joint state sequence through affine coupling. The dual-stream recursive unit sets up a steady-state recursive channel to capture the steady energy consumption evolution law and a perturbation recursive channel to capture the sudden perturbation response. The two channels are synthesized into the joint state sequence at each time step through a nonlinear fusion layer. The mask attention unit is used to receive the joint state sequence and the initial interval feature sequence, generate a context vector, and filter the attention weights by constructing a physical feasibility mask to shield features that do not meet the device's rated power or safety boundary, so that only time segment information that meets the physical constraints is retained after attention normalization, and the context vector is output. The constraint-embedded unit is used to receive the joint state sequence and the context vector, and combine the upper and lower limits of power, energy conservation constraints and climb rate constraints to generate condition codes. The constraint-embedded unit is implemented through condition vector construction and nonlinear projection, so that the condition codes can simultaneously embed statistical prediction laws and physical constraints, ensuring the consistency of subsequent generation results. The probability flow generation unit is used to receive the conditional coding, perform a layer-by-layer reversible transformation on the basic noise distribution and calculate the Jacobian determinant at each layer to obtain the probability prediction result of the energy consumption distribution. The probability flow generation unit is implemented through a reversible probability flow network and provides quantile prediction, confidence interval prediction and probability density function estimation results at the output end. The probability prediction results are used for subsequent energy consumption risk map construction and multi-level scheduling execution.

[0033] In this embodiment, the time-series decomposition unit first receives operating parameters, energy consumption curves, and environmental disturbance data. All data is aligned according to timestamps, ensuring that values ​​from different sensor terminals and data sources at the same time correspond to the same record line. After alignment, each continuous time series is divided into multiple scales: a shorter sliding window captures rapid fluctuations, a medium-length window characterizes daily periodicity, and a longer window reflects slow trends. The step size of the sliding window is smaller than the window length, allowing adjacent windows to partially overlap, thus avoiding excessive loss of information at boundaries. Wavelet decomposition uses a pair of grouped analytical filters to split the original sequence into low-frequency and high-frequency components. The low-frequency component is then decomposed repeatedly until a preset scale level is covered. In implementation, the sequence is passed sequentially through high-pass and low-pass filters to obtain a high-frequency system highlighting spikes and abrupt changes, and a low-frequency system representing slow trends, respectively. At each scale, the average value, the difference between the maximum and minimum values, the average change of adjacent samples, and the duration exceeding historical quantile levels are calculated within the window. These quantities are then concatenated chronologically to form an initial interval feature sequence. To facilitate the mixed use of different devices and different units of measurement, each feature is subtracted from the mean of the device over a recent period, and then divided by the standard deviation of the same period to obtain feature values ​​with uniform units of measurement and similar fluctuation ranges. These values ​​are then used as entries in the initial interval feature sequence and input into subsequent modules.

[0034] The dual-stream recursive unit receives the initial interval feature sequence and simultaneously maintains a steady-state hidden state and a perturbation hidden state for each time slice. The steady-state hidden state is used to accumulate the evolution law of steady energy consumption. Its update adopts a gated memory method: first, the initial interval feature sequence of the current time slice is compressed into a candidate vector; then, a retention weight between zero and one is calculated to represent the proportion of "retaining history", and another update weight between zero and one is calculated to represent the proportion of "accepting new information". The steady-state hidden state of the previous time step is multiplied by the retention weight, and then added to the candidate vector multiplied by the update weight to obtain the new steady-state hidden state. The perturbation hidden state is used to characterize the response to sudden perturbations. Its update process also adopts the same gating idea, but the generation of candidate vectors places more emphasis on fields such as high-frequency features, abrupt change magnitude, and duration. The steady-state hidden state and the perturbation hidden state are coupled affinely to obtain a joint state sequence. Specifically, the perturbation hidden state is mapped to a dimension-wise scaling factor and a dimension-wise translation. The scaling factor is restricted to a lower bound and an upper bound to ensure invertibility and stationarity. The steady-state hidden state is then stretched and translated dimension-wise using the scaling factor and the translation to generate the joint state sequence. At the same time, the joint state sequence is concatenated with the perturbation hidden state to serve as the complete temporal representation for subsequent attention calculations.

[0035] The masked attention unit receives a joint state sequence and an initial interval feature sequence. For each time slice, it constructs a vector representing the "query," a vector representing the "key," and a vector representing the "value." The relevance between the query and the key is obtained by summing the results after multiplying them dimension-wise. All scores are then shifted and scaled to bring them within a comparable range. To ensure attention is allocated only within physically feasible limits, a physical feasibility mask is introduced: any combination of time slices that would cause power to exceed the device's rated power, current to exceed the maximum allowable current, energy to exceed the energy conservation limit within a given window, or ramp rate to exceed a set limit, has its attention score set to zero. The remaining scores are normalized by summing all unmasked scores, resulting in a weighted average of one. The normalized weights are used to calculate a weighted average of the "value" vectors to obtain a context vector. This vector numerically emphasizes historical segments that are highly relevant to the current moment but do not violate device safety boundaries, thus incorporating both statistical relevance and physical constraints into the attention result.

[0036] The constrained chimera unit receives a joint state sequence and a context vector, aligns them in the time dimension, and concatenates them into an intermediate representation, introducing power upper and lower limits, energy conservation constraints, and ramp rate constraints. The power upper and lower limits are represented by normalizing the minimum and maximum available power for scheduling to a boundary vector between zero and one according to device type. The energy conservation constraint is obtained by accumulating the candidate power trajectories segment by segment at sampling time intervals within a given time window to obtain the energy accumulation, which is then encoded together with the normalized difference of the preset energy budget. The ramp rate constraint is obtained by calculating the power difference between adjacent time slices divided by the time interval to obtain the rate of change for each segment, which is then encoded together with the normalized difference of the maximum allowed rate of change. These constraints, along with the intermediate representation, are input into a nonlinear projector, which consists of several layers of bounded activation feedforward transforms. This projector compresses the intermediate representation into a fixed-length conditional encoding vector. To avoid the conditional coding from forcibly superimposing conflicting constraint information, the nonlinear projector calculates a set of coordination weights between zero and one before output. These weights redistribute the components according to the importance of the constraints, so that the conditional coding numerically takes into account both statistical prediction laws and physical constraints.

[0037] The probabilistic flow generation unit receives conditional encoding and starts with a set of easily sampled basic noise vectors. It performs reversible transformations layer by layer, progressively transforming the basic noise into energy-consuming samples that conform to the physically feasible domain. Each layer employs a coupled reversible structure: the vector is divided into two parts along a fixed dimension. The first part is used to perform several bounded activation feedforward operations to calculate the dimension-wise scaling factor and dimension-wise translation amount for the second part. The scaling factor is constrained between a non-zero lower bound and a finite upper bound to ensure reversibility. Subsequently, the second part is stretched and translated dimension by dimension using this scaling factor and translation amount. The next layer swaps the roles of the two parts to ensure that all dimensions are fully transformed. To assign the correct density to the samples in a probabilistic sense, each layer records the dimension-wise sum of the logarithmic absolute values ​​of its scaling factor. The results from each layer are then summed to obtain the total density correction amount. In this way, by accumulating scaling information layer by layer, the density change when mapping from basic noise to the target distribution can be obtained. To ensure that the generated energy consumption samples fall within the upper and lower power limits, each dimension is checked for out-of-bounds status immediately after each transformation. If an out-of-bounds sample is found, its scaling factor is reduced towards the boundary until the sample returns to within the boundary. Simultaneously, the corresponding boundary adjustment record is added to the density correction to maintain consistency between the probability density and the boundary treatment. Finally, a batch of energy consumption samples is obtained by repeating the sampling process multiple times and performing the same reversible transformation. These samples are then arranged in ascending order of value, and the position at a specified percentile is used as the quantile prediction for that percentile. The upper and lower bounds of a large concentrated region in the middle of the samples are used as confidence intervals. The estimated value of the probability density function is measured by the local density of the samples near a certain point. This local density can be obtained by setting a small numerical range around the point, counting the number of samples falling within that range, and dividing by the width of the range and the total number of samples. These statistics are combined with the density correction recorded at each generation to form the probabilistic prediction result of the energy consumption distribution.

[0038] The entire process executes sequentially from the output of the initial interval feature sequence by the temporal decomposition unit. The dual-stream recursive unit condenses the initial interval feature sequence into a joint state sequence. The masked attention unit calculates and outputs a context vector between the joint state sequence and the initial interval feature sequence. The constraint-based fusion unit combines the joint state sequence and the context vector with power upper and lower limits, energy conservation constraints, and ramp rate constraints to compress them into conditional codes. The probability flow generation unit receives the conditional codes and produces a probability prediction result of energy consumption distribution, including quantile prediction, confidence interval prediction, and probability density function estimation results. To ensure reproducibility, all gating weights and attention weights are obtained by linearly transforming the input values ​​to a controllable range and applying a monotonically bounded numerical compression function, ensuring that the weights always fall between zero and one. All normalization steps are performed by subtracting the recent mean and then dividing by the recent standard deviation. All threshold judgments are based on engineering parameters such as the equipment's rated power, maximum allowable current, energy budget, and maximum ramp rate, and the ratio between the detected quantity and the benchmark quantity is used to determine whether to mask or adjust.

[0039] Step S103: Apply an intervention signal within the equipment safety boundary, generate equipment response characteristics based on the intervention differences, and combine them with the probability prediction results to form an energy consumption risk map.

[0040] In step S103, intervention signals need to be applied to the equipment while ensuring its safe operation. This allows for observation of the equipment's response characteristics to minor external disturbances, thereby obtaining richer dynamic behavioral characteristics than simply relying on historical data. Intervention signals typically refer to slight adjustments to the equipment's operating parameters. The magnitude of these adjustments must be rigorously calculated and must not exceed the equipment's safety boundaries. For example, for motor-type equipment, short-duration pulse changes can be applied within 1% to 5% of the rated voltage or current; for refrigeration units, adjustments within 0.5°C of the control temperature setpoint can be applied; and for lighting loads, duty cycle changes for several seconds can be made within the brightness adjustment range. These signals must be reversible, short-duration, and precisely injected through the control system to avoid long-term impacts on the normal operation of the equipment.

[0041] After the intervention signal is applied, the changes in equipment operating parameters, energy consumption curves, and related environmental disturbance data before and after the intervention need to be recorded in real time using deployed distributed sensing terminals. To ensure the accuracy of the comparison, the triggering time of the intervention signal should be strictly aligned with the sensor sampling time, which can be achieved through the synchronization mechanism between the controller and the data acquisition system. By comparing the parameter differences before and after the intervention, the equipment's response vector under the intervention can be obtained, such as the transient fluctuation amplitude of the current, the adjustment trend of the power factor, or the short-term offset of the energy consumption curve. After these response vectors are accumulated through multiple interventions, a set of response characteristics of the equipment under different intervention conditions can be formed, thereby characterizing the equipment's sensitivity, inertia, and recovery capability.

[0042] The equipment response features are not used in isolation but need to be fused with the probability prediction results obtained in step S102. Specifically, the equipment response features can be used as additional feature channel inputs and cross-mapped with the probability prediction distribution to correct or supplement the prediction results. For example, if the probability prediction results indicate a high probability of significant energy consumption fluctuations in the future, and the equipment response features indicate that the equipment has a weak ability to recover from disturbances under specific interventions, then the risk level for this period should be significantly increased. Conversely, if the equipment demonstrates strong robustness in intervention tests, the corresponding risk level can be reduced accordingly.

[0043] Building upon the aforementioned integration, an energy consumption risk map is further constructed. This map uses time and equipment as coordinate dimensions, overlaying and mapping probabilistic prediction results with equipment response characteristics to form a risk matrix that changes over time. In the map, the intensity of color or numerical value indicates different levels of energy consumption anomaly risk; high-risk areas correspond to periods and equipment combinations where future energy consumption will exceed limits or fluctuate drastically, while low-risk areas correspond to periods and equipment combinations where operation is stable. The energy consumption risk map can be displayed as a two-dimensional heatmap or stored in a database as a matrix data structure for subsequent scheduling and decision-making.

[0044] Furthermore, the step of applying an intervention signal within the equipment safety boundary, generating equipment response characteristics based on the intervention differences, and combining them with probability prediction results to form an energy consumption risk map includes: Based on the equipment's rated power, maximum allowable current, temperature rise threshold, and ramp-up rate constraints, the equipment safety boundary is calculated, and the upper limit of the amplitude, duration, rise time, and spectrum limit of the intervention signal are determined accordingly. An intervention signal configuration table is then output. Based on the intervention signal configuration table, the current operating parameters, energy consumption curves, environmental disturbance data, and time synchronization benchmark, an unrelated intervention plan covering different time slices and different devices is generated. The unrelated intervention plan uses mutually orthogonal pseudo-random binary sequences and frequency sweep signals to avoid crosstalk, thereby obtaining the intervention plan and trigger timestamp set. An intervention signal is injected into the device control interface according to the set of trigger timestamps, and operating parameters, energy consumption curves and environmental disturbance data are collected synchronously within a fixed observation window before and after the intervention. Time alignment, detrending and bandpass filtering are performed to obtain the intervention window dataset. The intervention window dataset is used to estimate the response. The impulse response, gain, delay, time constant, recovery time and steady-state offset are calculated by a joint method of subspace identification and deconvolution. The confounding effects of environmental disturbance data on the response estimation are eliminated by instrumental variable regression to obtain the equipment response characteristics. The device response features are fused with the probability prediction results, and the distribution parameters of the probability prediction results are corrected based on likelihood weighting and Bayesian update to obtain the corrected probability prediction results. Using equipment and time slices as indexes, the out-of-limit probability, variance, and tail thickness of the corrected probability prediction results are compared with the equipment safety boundary to generate map entries containing risk intensity, source tags, and time location, and aggregated according to the index to form an energy consumption risk map.

[0045] In this embodiment, the rated power, maximum allowable current, temperature rise threshold, and ramp-up rate constraints are first converted to the same measurement system of "power offset - duration - change slope" to facilitate comparison and selection of the most stringent boundaries within the same table. Rated power and maximum allowable current jointly limit the upper limit of allowable power offset: based on the current operating point, the increased power must not cause the instantaneous power to exceed the rated power, nor should it cause the converted current to exceed the maximum allowable current; to avoid exceeding the limits, the smaller margin given by the two is taken as the upper limit of the amplitude. The temperature rise threshold limits the allowable energy injection: first, the temperature rise caused by unit energy is obtained using the historical temperature rise-energy correspondence; then, the temperature rise threshold is converted into an injectable energy budget, requiring that "power offset multiplied by intervention duration" does not exceed this budget, thus obtaining the upper limit of duration. The ramp-up rate constraint limits the minimum time for the rise edge: the allowable power change per unit time is compared with the target amplitude to obtain the shortest time required to smoothly increase the power from the current value to the upper limit of the amplitude, which is taken as the lower limit of the rise edge. To avoid triggering inherent resonance in the equipment or sensitive frequency bands of the power grid, a list of prohibited frequency ranges was compiled based on previous frequency response test records, stipulating that the spectrum of the intervention signal must not include these ranges. The above-mentioned upper limit of amplitude, upper limit of duration, lower limit of rise time, and spectrum restrictions, along with equipment identification, cooling interval, and maximum number of interventions per day, are recorded together to form an "Intervention Signal Configuration Table".

[0046] After obtaining the "Intervention Signal Configuration Table," "Unrelated Intervention Plans" are generated for multiple devices and multiple time slices. To avoid crosstalk between devices and interference between adjacent interventions on the same device, each device is assigned a pseudo-random binary sequence that is approximately orthogonal to other devices within the observation window. Orthogonality means that the product of any two sequences at the same position is close to zero after the entire sequence is accumulated, so their mutual influence can be statistically canceled out. For devices requiring frequency domain identification, additional frequency sweep signals are added, but different devices are assigned non-overlapping frequency sweep coverage areas, and the start and end frequencies of the frequency sweeps avoid the prohibited frequency bands listed in the "Intervention Signal Configuration Table." Using a unified time synchronization reference, precise start and end times are specified for each sequence and each frequency sweep, and a window period of not less than the cooling interval is maintained between adjacent interventions, thus obtaining the "Intervention Plan and Trigger Timestamp Set."

[0047] According to the "Intervention Plan and Trigger Timestamp Set," intervention signals are injected at the specified trigger time via the device control interface. To ensure the comparability of subsequent comparative analyses, a fixed observation window is set around the trigger time for each intervention. The window includes the pre-intervention baseline interval and the post-intervention response interval. The length of the baseline interval is taken as several times the known dominant time constant to stably estimate the baseline mean and slow drift; the length of the response interval is taken as the upper limit of the time required for the output to recover to a small range of fluctuations near the baseline. Operating parameters, energy consumption curves, and environmental disturbance data are recorded simultaneously within the window. All data are first aligned with the trigger time. If there is a sampling phase difference, data with different sampling phases are aligned to the same time scale through interpolation. Then, detrending is performed on each channel, and the slow slope in the baseline interval is estimated using local linear fitting and subtracted from the entire window to remove slow drift. Then, bandpass filtering is applied to retain the components within the frequency band allowed by the "Intervention Signal Configuration Table" and suppress low-frequency slow changes and high-frequency noise to obtain the "Intervention Window Dataset."

[0048] The "intervention window dataset" is used for response estimation. To extract the dynamic characteristics of the system, the intervention input and corresponding output are first organized into overlapping blocks in time. The blocks are adjacent to each other, and the original order of the input and output is preserved within each block, thus forming stacked data that can characterize the evolution of the system state. By performing orthogonal decomposition on the predictable and unpredictable parts between these overlapping blocks, the low-dimensional state trajectory that best explains the output change is separated, and the main dynamics of the system under undisturbed and disturbed conditions are obtained. Since the intervention input is known, the spectrum of the input is used as a "template" in the frequency domain. The output component at the same frequency point is divided by the amplitude of the template and the phase is corrected to obtain the output change per unit input. This is then summarized along the frequency axis and restored to the time axis to obtain the "impulse response". In the time domain, to reduce the confounding caused by environmental disturbances, environmental disturbance data is used as instrumental variables in the regression: disturbance channels that are almost uncorrelated with the intervention input but affect the output are selected as instrumental variables. The output is first predicted using these instrumental variables, and then the prediction residuals are regressed against the intervention input so that the regression coefficients only reflect the direct effect of the intervention input on the output. The "gain" is obtained from the "impulse response," and is calculated by finding the average change in output after stabilization within the response interval, and expressing it as the ratio of this value to the amplitude of the intervention input. "Time delay" is the time from the trigger moment to the first significant deviation of the output from the baseline; the deviation threshold can be taken as a small fraction of the stable change. "Time constant" is the time it takes for the output to reach approximately two-thirds of its stable change during the rise process. "Recovery time" is the time it takes for the output to return to a small range of its stable change after the intervention ends. "Steady-state offset" is the residual difference between the output and the baseline a considerable time after the intervention ends. These values, along with their uncertainties, constitute the "device response characteristics."

[0049] When fusing "device response characteristics" with "probabilistic prediction results," the "probabilistic prediction results" are considered a priori description of future energy consumption. This typically includes information characterizing the central location and dispersion, as well as information related to high-energy-consumption tails. The credibility of the prior is evaluated based on the "device response characteristics": if the average change predicted by the prior matches the "gain" and "steady-state offset," and the prior uncertainty covers the typical range of "time constant" and "delay," then the prior scenario is considered more credible and given a higher weight; if the prior deviates significantly from the observed characteristics, a lower weight is given. The weights are calculated by normalizing each deviation using its allowable range, then taking a weighted average of the normalized deviations and mapping it back to a score between zero and one. Smaller deviations result in higher scores, and these scores are used as the weights for the scenario. The parameters of all scenarios are weighted and summed to obtain the center position of the "corrected probability prediction result". The dispersion of each scenario from the center is also weighted and summed to obtain the dispersion of the "corrected probability prediction result". For tail-related indicators, the high tail quality of scenarios with larger weights that are consistent with "gain" and "steady-state offset" is retained, while the contribution of scenarios with smaller weights to the tail is compressed. In this way, the correction based on the idea of ​​likelihood weighting and Bayesian update is completed.

[0050] Using equipment and time slices as indices, entries for mapping are obtained by comparing each item against the "corrected probability prediction results" and "equipment safety boundaries." The probability of exceeding limits is calculated as follows: in the distribution described by the "corrected probability prediction results," the proportion of probability quality above the "equipment safety boundaries" is counted; variance is obtained by averaging the deviations of each possible value from the center position according to its probability magnitude; tail thickness is characterized by comparing the probability quality at the high quantile to that in the middle region—the larger the ratio, the thicker the tail. The probability of exceeding limits, variance, and tail thickness are normalized to a uniform zero-to-one scale and then synthesized into "risk intensity." Higher weight can be given to the probability of exceeding limits during synthesis to reflect the priority of risk exceeding the boundary; "source label" identifies the main source of risk—whether it is caused by intervention response sensitivity, prediction uncertainty, or environmental disturbances—by comparing the contributions of the three types of factors to the changes in the "corrected probability prediction results," with the largest contribution being the label; "time location" is the specific time slice corresponding to the entry. By continuously stitching together the map entries of the same device in adjacent time slices, and juxtaposing and aggregating the entries of different devices in the same time slice, an indexable, searchable, and visualized "energy consumption risk map" is formed. It is stored with the same time base as the "intervention signal configuration table" and the "intervention plan and trigger timestamp set" to ensure that subsequent digital twin simulations and multi-level scheduling can be seamlessly called.

[0051] Step S104: Based on the energy consumption risk map, generate extreme disturbance scenarios using digital twins, and verify the consistency between the probability prediction results and the scheduling scheme in the data domain and physical domain. Perform multi-level scheduling based on the verified energy consumption risk map and task delayability to form the scheduling result.

[0052] In step S104, the energy consumption risk map generated in the previous step is used as input. High-risk and potential anomaly areas identified in this map are used to drive the construction of a virtual operating environment using a digital twin model. The digital twin model should be a highly realistic simulation of the actual equipment and system operation, with parameters including electrical characteristics, equipment performance curves, environmental disturbance models, and operating constraints. This model is initialized by aligning with actual collected data to accurately reflect the current system operating state. During the simulation, extreme disturbance scenarios need to be artificially injected, such as simulating sudden load increases, drastic changes in ambient temperature, fluctuations in power supply voltage, or simultaneous start-up and shutdown of multiple devices, to verify the system's energy consumption performance and stability under the most unfavorable conditions. The generation of these disturbance scenarios should be based on the risk level distribution in the risk map to ensure that the simulation covers extreme situations that may occur in actual operation.

[0053] After generating extreme disturbance scenarios using a digital twin, the energy consumption response results output by the model need to be compared one by one with the probabilistic prediction results in step S102. During the data domain verification process, the focus is on checking whether the distribution of the probabilistic prediction results is consistent with the simulation results in terms of statistical characteristics, including whether the mean, variance, and distribution interval are within a reasonable error range. If there is a significant deviation, such as the prediction results being overly optimistic and not covering the extreme energy consumption values ​​that occur in the simulation, the prediction results should be deemed unreliable. During the physical domain verification process, it is necessary to verify whether the scheduling scheme meets the physical constraints of the equipment and system, such as the rated power of the equipment, maximum allowable current, temperature rise limit, and power supply network capacity. If the scheduling scheme leads to operating results exceeding the physical boundary conditions in the simulation scenario, the scheduling scheme should be deemed infeasible.

[0054] By completing the above dual verification, prediction results and scheduling schemes that conform to both statistical distribution rationality and physical constraints can be selected. Based on this, it is necessary to further integrate task delayability information to classify different tasks according to their urgency and time flexibility. For tasks that can be delayed, execution can be appropriately postponed during high-risk periods; for tasks that must be executed immediately, redundant resources need to be reserved in the scheduling. The final multi-level scheduling process is optimized based on the verified energy consumption risk map and task classification results, generating scheduling results that both ensure safe equipment operation and improve overall energy efficiency.

[0055] This process ensures that the scheduling scheme does not rely solely on model predictions, but is verified through extreme simulations using digital twins and dual checks in the data and physical domains, making the final scheduling results more reliable and feasible.

[0056] Furthermore, the process of generating extreme disturbance scenarios using digital twins based on energy consumption risk maps, and verifying the consistency between probability prediction results and scheduling schemes in the data and physical domains, and performing multi-level scheduling based on the verified energy consumption risk maps and task delayability to form scheduling results, includes: Based on the energy consumption risk map, high-risk areas and potential abnormal areas are extracted and marked. A risk triggering condition set is generated by combining operating parameters, energy consumption curves and environmental disturbance data. A digital twin simulation environment is constructed based on the risk triggering condition set. The physical characteristics of the equipment, energy consumption curves and environmental disturbance models are loaded into the simulation environment, and the marked areas are initialized to obtain the initial state of the simulation. Extreme disturbance trajectories are injected into the initial state of the simulation. These extreme disturbance trajectories include scenarios such as power surge, temperature change, voltage fluctuation, and concurrent switching of multiple devices, generating a corresponding simulation response dataset. The simulation response dataset and the probability prediction result are compared in the data domain, the mean, variance and quantile deviation are calculated, and the probability prediction result is corrected when the deviation exceeds a set threshold to obtain the probability prediction result corrected by the data domain. The probability prediction results corrected by the data domain are verified against the scheduling scheme in the physical domain to check whether the scheduling scheme meets the constraints of equipment rated power, maximum allowable current, energy conservation and ramp rate, and to obtain the scheduling scheme verified by the physical domain. By combining the probability prediction results corrected by the data domain with the scheduling scheme verified by the physical domain, the energy consumption risk map is updated and the task delayability is marked. Multi-level scheduling is then performed according to the task delayability to obtain the confirmed scheduling results.

[0057] In this implementation, the target area to be simulated is first extracted based on the energy consumption risk map. The risk intensity of all time slices and equipment combinations within one observation period in the energy consumption risk map is extracted and arranged in ascending order. The boundary value at which 80% of the samples are no higher than the risk intensity, while 20% are higher, is identified. Items exceeding this boundary value are marked as high-risk areas. Items within 10% above and below this boundary value, whose risk intensity has not significantly decreased after the most recent intervention, are marked as potential anomaly areas. For each marked item, the operating parameters, energy consumption curves, and environmental disturbance data for the same time slice are further read, and the center position, dispersion, and several high quantile values ​​of the corresponding time slice in the "probability prediction result" are recorded. This information is merged into a risk trigger condition set. A record in the condition set consists of the equipment identifier, time slice identifier, risk intensity, corresponding operating parameters, the current value and recent change range of the energy consumption curve and environmental disturbance data, and the center position and key quantile values ​​of the probability prediction result.

[0058] When constructing a digital twin simulation environment based on the risk trigger condition set, it is necessary to load the physical characteristics of the equipment, energy consumption curves, and environmental disturbance models, and initialize the states corresponding to the marked areas. The physical characteristics of the equipment are derived from the equipment nameplate and acceptance records, including rated power, maximum allowable current, heat capacity parameters, heat transfer capacity, allowable ramp-up rate, and efficiency curves. If any parameters are missing, they are calibrated using data from historical steady-state operating periods: within a time period without intervention, the power and temperature rise within an equal-length window are taken, and the temperature rise caused by a unit of energy is calculated. This is achieved by adding the power within the window over time to obtain the energy, and then dividing this energy by the temperature rise within the same window to obtain a conversion factor. This factor is then used to backfill the heat capacity or heat transfer parameters. The environmental disturbance model is constructed based on the environmental disturbance data recorded in the risk trigger condition set: the background trajectory is obtained by extrapolating the recent trends in temperature, humidity, and airflow; occasional events such as access control opening and material loading are implanted according to their trigger times. The energy consumption curve is initialized using the most recently available baseline segment, and the initial value of the baseline is corrected using the current operating parameters at the initial moment of the simulation to align the digital twin model with the actual field conditions. After alignment is completed, the above parameters and states are used as the initial state for simulation.

[0059] When injecting extreme disturbance trajectories in the initial simulation state, power surges, temperature jumps, voltage fluctuations, and concurrent switching of multiple devices need to be converted into executable time series. The magnitude of a power surge must not exceed the upper limit of the amplitude in the device's safety boundary, the duration must not exceed the upper limit of the duration, and the rise time must not be less than the lower limit of the rise time. During generation, the current power is used as the starting point, a linear ramp-up segment raises the power to the target amplitude, maintains that level until the duration ends, and finally a linear fall-down segment lowers the power back to the baseline. Temperature jumps are achieved by raising or lowering the ambient temperature trajectory by a specific amplitude within a short period. The amplitude is taken from historical extreme values ​​or the upper limit of external temperature change given by the safety assessment. The ramp-up or fall-down time is not less than the time constant corresponding to the device's thermal inertia to avoid unrealistic instantaneous jumps. Voltage fluctuations are achieved by setting symmetrical offsets above and below the rated voltage and applying them in a sinusoidal or step manner over a short period, with the maximum offset not exceeding the upper limit of the power supply's allowable voltage deviation. Concurrent switching of multiple devices is achieved by simultaneously starting, stopping, or switching gears on multiple devices within the same time slice. However, to avoid grid impact, small peak shifting steps on the order of milliseconds to seconds are set according to power from highest to lowest, ensuring that the overall combined ramp-up rate does not exceed the station-level allowable limit. For each extreme disturbance trajectory, the power, temperature rise, and key operating parameters of the affected devices are recorded over time in the simulation. The outputs from each trajectory are then stitched together to form a simulation response dataset. To facilitate subsequent statistical analysis, the simulation response dataset saves a set of multiple trajectories for each time slice to approximate various extreme scenarios that may occur in the field.

[0060] When comparing the distribution of the simulated response dataset and the probabilistic prediction results in the data domain, it is necessary to provide the calculation methods for the mean, variance, and quantile deviation. The mean is obtained by adding all simulated values ​​for the same time slice and the same index (e.g., power), and then dividing by the number of simulated trajectories. The variance is obtained by first calculating the mean for that time slice, then calculating the difference between each simulated value and the mean, summing the squares of these differences, and then dividing by the number of trajectories. The quantile is obtained by sorting all simulated values ​​for the same time slice from smallest to largest, and taking the value that falls at the target percentage. For example, to take the 95th percentile, locate the position near the top after sorting and read its value. During the comparison, the mean deviation is obtained by subtracting the center position of the simulated mean from the center position of the probabilistic prediction results; the variance deviation is obtained by subtracting the dispersion of the simulated variance from the dispersion of the probabilistic prediction results; and the quantile deviation is obtained by subtracting the quantile of the simulated quantile from the corresponding quantile value of the probabilistic prediction results. The threshold setting employs a historical comparison quantile upper bound method: During periods without past anomalies, the comparison process is repeated to obtain a set of historical deviation values. These historical deviations are then sorted from smallest to largest, and the value closest to the top of the sorted list is taken as the upper limit of the allowable bandwidth. Specifically, this value can be chosen such that only a small portion of the historical deviations exceed this limit. If any current deviation exceeds its allowable bandwidth upper limit, the probability prediction result is deemed to require correction. During correction, the center position of the probability prediction result is first shifted towards the simulation mean by an amount equal to the mean deviation. Then, the dispersion of the probability prediction result is amplified or reduced proportionally to the simulation variance and the original dispersion; that is, it is amplified when the simulation variance is larger and contracted when it is smaller. For quantiles, the higher and lower quantiles are extrapolated or inverted in the same direction and proportion, preserving the asymmetric characteristics consistent with the device response direction. After this shifting and scaling, the data domain-corrected probability prediction result is obtained.

[0061] The probability prediction results corrected for the data domain are validated against the scheduling scheme in the physical domain. This requires checking the power settings of each device and the derived current within the device's safety boundaries for each time slot. Current calculation is based on the relationship between voltage and power: given the voltage level and power factor, the power is divided by the product of voltage and power factor to obtain an estimated current value, which is then compared with the maximum allowable current. If the voltage fluctuates within a time slot, the same conversion is performed using the voltage value for that time slot. Energy conservation is verified by multiplying the power of each time slot by its duration within a preset time window to obtain the corresponding energy. The energy of all time slots within the window is then summed to obtain the total energy. This total energy is compared with the allowable energy budget for that window; if it exceeds the allowable budget, the scheduling scheme is considered unqualified. The climb rate is verified by subtracting the power settings of two adjacent time slots and dividing by the time interval between them to obtain the rate of change. This rate is then compared with the maximum allowable climb rate. The same method is used for station-level constraints, but the sum of device power is used as the verification object. All non-compliant entries are marked and sent back to the scheduler. The scheduler must make minimal adjustments to the marked entries without altering the already approved parts, prioritizing reducing constraints or moving out-of-bounds tasks to adjacent low-risk time slices, until all constraints are met or the task is determined to be unable to complete within the current window. After completing the above process, a scheduling scheme verified by the physical domain is obtained.

[0062] When using the data domain-corrected probability prediction results and the physical domain-verified scheduling scheme to update the energy consumption risk map and label task delayability, the original distribution parameters and risk intensity are first replaced under the same index in the risk map. The risk intensity is recalculated by weighting the normalized values ​​of the over-limit probability, variance, and tail thickness. The over-limit probability related to exceeding the limit is given the highest weight, followed by dispersion, and then tail thickness. The sum of the weights is one, and the composite value is the new risk intensity. The labeling of task delayability depends on the earliest start time, latest finish time, duration, and power requirement of the task. First, all time slices that the task can cover within the available window are listed. Time slices with risk intensity exceeding a certain upper limit and no feasible verification scheme are removed. Then, the length of the longest continuous interval that can be coherently arranged is calculated in the remaining time slices. If this length is not less than the task duration, the task is labeled as delayable, and its set of placement intervals is marked. If it is insufficient, it is labeled as non-delayable or needs to be split for execution. Multi-level scheduling proceeds hierarchically: The first level, within the equipment layer, moves deferred tasks from time slots with higher risk intensity to those with lower risk intensity. The second level, at the equipment group level, redistributes tasks among similar equipment to flatten peak power within the group, while preserving the low-risk arrangement established in the first level. The third level, at the station level, checks the power of all equipment. If any exceedances of the station-level ramp-up rate or energy budget still exist, fine-tuning is performed on edge time slots based on the principle of minimizing operational impact, prioritizing time slots with the lowest risk intensity to ensure tasks are not pushed back to high-risk periods. After each adjustment, the probability prediction results, corrected by the data domain, are immediately used to quickly estimate the change in the probability of exceeding limits, and the corresponding entries in the risk map are updated synchronously until all constraints are met and no further adjustments can reduce the risk intensity or station-level exceedances occur, thus obtaining a confirmed scheduling result.

[0063] Through the above process, starting from the energy consumption risk map, a risk triggering condition set can be objectively and repeatably constructed, a simulation initial state consistent with the field can be established, extreme disturbance trajectories covering key adverse situations can be generated, distribution comparison and correction based on mean, variance and quantile can be completed, physical constraint verification can be performed at the time slice, device and station levels, and the obtained probability prediction results corrected by the data domain and the scheduling scheme verified by the physical domain can be written back to the risk map. Multi-level scheduling can be completed under clear task delay labeling, and finally a confirmed scheduling result that can be directly issued can be formed.

[0064] Step S105: Send the scheduling results to the device and send back feedback data, and update the machine learning model based on the feedback data.

[0065] In step S105, the scheduling results generated in the previous step through consistency verification and multi-level scheduling need to be distributed to each target device in the form of explicit control commands. To ensure that the commands can be accurately recognized and executed by the devices, the scheduling results should be converted into a command set conforming to the device communication protocol, such as Modbus, BACnet, MQTT, or a custom industrial control message format. These commands not only include the device's switching control information but also power setpoints, operating periods, load allocation ratios, and necessary timestamp information, so that the devices can operate according to the scheduling requirements within the specified time window. During the distribution process, commands can be broadcast uniformly through a centralized control server, or a distributed control strategy can be adopted, with edge nodes sending commands to the devices under their jurisdiction one by one according to the scheduling results. To prevent commands from being tampered with or lost during transmission, the communication link needs to be overlaid with encryption and verification mechanisms, such as attaching checksums, digital signatures, or two-way handshake confirmation mechanisms to the messages, thereby ensuring the integrity and reliability of the scheduling commands.

[0066] When equipment receives and executes a scheduling instruction, it needs to transmit feedback data in real time to confirm the execution status. Feedback data includes the equipment's actual operating parameters, execution result status, energy consumption curve changes, and any abnormal information. For example, if a scheduling instruction requires a cooling unit to reduce its operating power by 10%, the feedback data should reflect whether the unit's actual output power has decreased as required, and simultaneously record the energy consumption trend. If the equipment fails to execute the instruction due to physical limitations or a sudden malfunction, an error code or abnormal status indicator should be included in the feedback. The frequency of feedback data acquisition needs to match the scheduling time precision; it can be set to second-level or minute-level sampling and transmitted back to the central control system via the same transmission link. All feedback data should be uniformly stored in a time-series database and timestamped with the issued scheduling instructions to ensure a one-to-one correspondence between execution and feedback.

[0067] After collecting feedback data, this data needs to be used as new input to update the machine learning model. Updates can be achieved through online learning or incremental training, where feedback data is continuously added as new samples to adjust the model's parameter weights without discarding existing training parameters. To avoid overfitting the model to short-term fluctuating data, the feedback data can be smoothed or a sliding time window can be used, combining the most recent data with historical data. During the update process, the loss function can be recalculated, the difference between the model's predictions and the feedback data can be compared, and parameters can be adjusted using the backpropagation algorithm. In this way, the model can continuously absorb new data during operation, adapting to changes in equipment status and environmental conditions, thereby maintaining the accuracy of prediction and scheduling.

[0068] A continuous adaptive optimization mechanism can be formed through a closed-loop process of issuing scheduling results, collecting feedback data, and updating the model. The issuance of scheduling instructions ensures that predictions and decisions are truly implemented at the device level, the feedback data ensures that the execution status is monitored and recorded in real time, and the model update ensures that future predictions are more consistent with the actual operating conditions.

[0069] Furthermore, the step of sending the scheduling results to the device and transmitting feedback data back, and updating the machine learning model based on the feedback data, includes: The confirmed scheduling results are converted into a set of energy consumption management instructions that can be recognized by the device. The set of energy consumption management instructions includes power settings, task execution time windows, load allocation ratios and safety redundancy parameters. It is then sent to the target device through an encrypted communication protocol with digital signatures and timestamps to obtain a record of the sent instructions with a unique identifier. On the device side, energy consumption control operations are performed according to the issued instructions. During the execution process, operating parameters, energy consumption curves and environmental disturbance data are collected in real time to generate an execution feedback stream with an instruction identifier. The execution feedback stream is transmitted back to the data aggregation node through the same encrypted communication protocol to obtain a feedback dataset bound by the identifier. At the data aggregation node, the feedback dataset and the issued instruction record are timestamped and their integrity is verified. If the verification result shows that the feedback dataset and the issued instruction record are consistent, the feedback dataset is stored as a trusted execution sample. During the model update phase, the trusted execution sample library is fused with the historical training set of the machine learning model, and the execution data of the most recent time period is selected using the sliding window method to construct a dynamic training subset. Incremental training is performed on the dynamic training subset in the machine learning model. The model parameters are corrected by minimizing the probability distribution difference and Bayesian regularization to generate an online updated machine learning model. The online updated machine learning model is then used for the probability prediction of subsequent energy consumption distribution to obtain a closed-loop adaptive prediction model.

[0070] In this embodiment, the confirmed scheduling results need to be transformed into a set of energy-managed instructions that the devices can execute unambiguously, and it is ensured that instructions and feedback can be reliably associated and verified throughout the entire process of network transmission and device execution. Specifically, each plan in the scheduling results is broken down into clearly defined fields, including device identifier, power setting value, task execution time window, load allocation ratio, and safety redundancy parameter, with unified units and value ranges. The device identifier uses a string corresponding to the device's serial number; the power setting value is expressed as a percentage of rated power; the task execution time window is expressed as the start and end times in Coordinated Universal Time (UTC); the load allocation ratio is expressed as a decimal between zero and one, with the sum of the ratios of all parallel tasks not exceeding one; and the safety redundancy parameter is expressed as a percentage downward offset of the power setting value, used for automatic downgrading in case of anomalies. These fields are written into the payload area of ​​the binary frame in a fixed order. The frame header contains the protocol version and device identifier, and the frame tail reserves a signature area. To assign a unique identifier to each issued instruction record, an irreversible digest calculation is performed on the frame header and payload area to obtain a fixed-length digest value. This digest is then concatenated with a millisecond-level timestamp, scheduling batch number, and sequence number. A digital signature is generated using the same private key, and the timestamp and digital signature are written to the frame tail. Simultaneously, a unique instruction number is constructed by concatenating the textual representation of the digest value with the timestamp. This instruction number, device identifier, and original field content are registered in the control terminal database, forming the issued instruction record. The encrypted communication protocol employs a two-way authentication and transmission encryption capability. When establishing a connection, the device and control terminal mutually verify each other's certificates. During transmission, all byte streams are encrypted and protected for integrity at the link layer; any tampering during transmission will be rejected at the receiving end.

[0071] After receiving the energy consumption managed instruction set, the device first verifies whether the signature was issued by the trusted control terminal using its built-in public key, and checks whether the difference between the timestamp and the local clock is within the allowed time difference range. If it exceeds the range, the instruction is rejected and an error status is returned. After successful verification, the device writes the instruction number and field content into the execution queue, inserts or overwrites the timing table of the local scheduler according to the task execution time window, and performs boundary checks on the power setpoint and load allocation ratio to ensure that they do not exceed the device's security boundary and do not conflict with the current operating mode. During execution, the device collects operating parameters, energy consumption curves, and environmental disturbance data in real time at a fixed sampling period; each batch of sampling points is bound to the instruction number and the current sub-time period identifier to ensure one-to-one correspondence with the instruction. To avoid packet loss caused by network fluctuations, the device caches the sampling points in a circular cache in chronological order and sets a minimum batch threshold or a maximum waiting time. If either condition is met, the sample is packaged into an execution feedback stream and reported upwards. The data blocks of the execution feedback stream are transmitted back using the same encrypted communication protocol as those sent. The header of the data block carries the device identifier and instruction number, and the payload area includes the sampling timestamp, real-time power, real-time voltage and current, real-time cumulative energy consumption, ambient temperature and humidity, and execution status code (such as normal, downgrade, and rejection). The tail includes the device's signature and verification fields. After receiving the execution feedback stream, the data aggregation node at the control end first verifies the device signature and link integrity, and then timestamps the feedback data and the sent instruction record using a unified time base in milliseconds. The alignment method is to calculate the clock deviation between the two ends and use this deviation to shift the entire feedback timestamp so that the trigger time coincides with the planned start time. If there are a few duplicate timestamps, the last one to arrive is used. If there are short-term missing timestamps, the values ​​of two adjacent sampling points are used for linear interpolation to fill in the gaps. Subsequently, each data block undergoes integrity verification, checking whether the instruction number, device identifier, and issued instruction record are consistent, whether the sub-time period falls entirely within the task execution time window, and whether the sampling sequence is continuous without breaks. If all checks pass, the data block is marked as trustworthy and written to the time-series storage bound to the instruction number. When all execution time periods corresponding to a certain instruction number are covered by trustworthy data, all feedback for that instruction is merged into a trustworthy execution sample. All trustworthy execution samples are archived in chronological order and by device dimension, accumulating to form a trustworthy execution sample library.

[0072] To ensure that the samples updated in the model represent the latest operating conditions without overly relying on short-term randomness, a dynamic training subset is constructed by fusing a trusted execution sample library with historical training sets. First, a sliding window length and sliding step size are determined. The window length can be set according to the rate of change of the target scenario's operating conditions; for example, a data center might be able to use the data for one to two weeks, while heavy industrial production might only be able to use it for a few days. All trusted execution samples covered by the last complete window on the timeline are selected into the candidate set. To avoid poor-quality data affecting the model, the quality of each sample is scored in two parts: one part reflects the completeness of the data, scored according to the proportion of valid sampling points to the planned number of sampling points (higher proportion, higher score); the other part reflects the fit of the execution, scored according to the proportion of time within the sampling segment where the "deviation between actual power and power setpoint" falls within the allowable range (higher proportion, higher score). The two scores are added together with equal weight to obtain the quality score, and samples with quality scores below a threshold are removed from the candidate set. The retained samples are assigned time-decay weights, with those closer to the current sample having a larger weight and those farther away having a smaller weight. The weights are calculated by queuing the samples within the window in chronological order, setting the weight of the closest sample as the baseline (1), and then multiplying the weight by a decay coefficient less than one for each subsequent equal interval until the beginning of the window. The quality score of each sample is multiplied by its time weight to obtain the final sample weight. Based on this weight, several appropriately sized mini-batches of data are extracted from the candidate set to form a dynamic training subset. Each mini-batch maintains diversity in chronological order and device distribution to improve the stability of updates.

[0073] When performing incremental training in machine learning models, it is necessary to combine probability distribution difference minimization and Bayesian regularization to refine model parameters. The goal of probability distribution difference minimization is to make the energy consumption distribution output by the model more closely resemble the empirical distribution reflected in reliable execution samples. Specifically, for each time slice and device, actual power samples for that time slice are collected in a dynamic training subset. These samples are arranged from smallest to largest, and values ​​at several fixed percentile positions are taken as empirical quantiles. Simultaneously, the current model outputs the corresponding predicted quantile for the same time slice. The difference between each pair of empirical and predicted quantiles is calculated; a positive difference indicates the model underestimates, and a negative difference indicates the model overestimates. During parameter updates, model weights are adjusted in directions that reduce these differences: if the model underestimates, the output related to that time slice is increased; if it overestimates, it is decreased. The magnitude of the adjustment is proportional to the absolute value of the difference and multiplied by the final sample weight, ensuring that high-quality, recent samples have a greater impact on the update. To avoid instability caused by over-updating, Bayesian regularization is incorporated, using the parameters of the "machine learning model updated in the last iteration" as a reference center. Parameters that deviate too much are pulled back. Specifically, after each round of parameter adjustment, the change in each parameter relative to the reference center is calculated, and then compared with the parameter's stability in historical training: parameters with more stable history are allowed smaller changes; if the actual change exceeds the allowable range, the adjustment amount for that parameter in this round is reduced proportionally. Parameters with historically large fluctuations are given slightly higher allowable ranges to better adapt to new conditions. Each round of updates undergoes rapid testing on a validation slice composed of recent samples. The testing criteria include whether the average difference between empirical quantiles and predicted quantiles has decreased, whether the out-of-limit probability approaches the actual occurrence ratio of reliable execution samples, and whether physical consistency is still satisfied after the injection of constraints. If any of these deteriorates, the system immediately rolls back to the previous round of parameters and reduces the adjustment step size before trying again.

[0074] To ensure the security and traceability of online deployment, all incremental training processes are version-managed. A version number is generated for each parameter update, which is a concatenation of the date and time, the start and end of the data window, and a summary of the dynamic training subset. This version number, along with the configuration, metrics, and validation results used during training, is written into the model registry. The online model is only replaced when the metrics of the online-updated machine learning model on the validation slice reach the preset thresholds. During replacement, the old version is retained as a backup, allowing for rapid rollback in case of anomalies in the new version. After deployment, an output calibration is performed using a two-step method: temperature scaling and quantile alignment. Temperature scaling globally stretches or compresses the dispersion of the model output to ensure that the coverage on the recent validation slice matches the target confidence level. Quantile alignment reduces the gap between the predicted and validation quantiles to within a set tolerance band at multiple fixed percentiles. This is achieved by adjusting the output in small increments around the data near these percentiles, ensuring that the predictions align with empirical data at these key points while maintaining the overall order.

[0075] After completing the above steps, the online-updated machine learning model is reconnected to the prediction-scheduling link and marked as a closed-loop adaptive prediction model. The maintenance of the closed loop relies on two types of continuous monitoring: one is data drift and execution quality monitoring, which summarizes the distribution of the difference between the empirical quantile and the predicted quantile on an hourly basis, calculates the difference between the actual and predicted proportions of exceeding limits on a daily basis, and calculates the integrity score and execution fit score of the feedback data in real time; the other is stability monitoring, which calculates the cumulative magnitude of parameter changes after multiple consecutive updates. If the cumulative magnitude exceeds a pre-set upper limit, a cold start retraining or an extension of the window to reduce the update frequency is triggered. When any monitoring item exceeds the threshold, the system automatically reduces the incremental training step size, increases the length of the validation slice, and rolls back to the previous stable version if necessary. Through this closed-loop chain of "instruction generation, encrypted distribution, identifier binding, time alignment, integrity verification, sample storage, window extraction, weighted incremental processing, regularization constraints, version release, online calibration, and continuous monitoring," the scheduling results are reliably delivered to the devices, the execution feedback stream is seamlessly transmitted back to the data aggregation node, and the trusted execution sample library and dynamic training subset provide the model with continuously fresh and usable information. The online-updated machine learning model can improve the probability prediction results in a verifiable manner after each round of updates, thus achieving steady-state operation as a closed-loop adaptive prediction model.

[0076] Furthermore, the step of fusing the trusted execution sample library with the historical training set of the machine learning model during the model update phase, and using a sliding window method to select the execution data from the most recent time period to construct a dynamic training subset, includes: Each sample in the trusted execution sample library is indexed based on timestamp and device identifier, and power execution deviation, energy consumption accumulation deviation and environmental response deviation are calculated within the same time slice. The three are combined into a quality score vector to obtain a sample index table with quality scores. According to the sample index table, the baseline samples corresponding to the time slice and device category are retrieved in the historical training set, and the trusted execution samples are paired and fused with the baseline samples using the quality score vector as the weight to generate a fused sample sequence with a bias correction factor. A sliding window is constructed on the fused sample sequence. The window size is set according to the energy consumption fluctuation cycle of the target scene. The window step size is smaller than the window size to form an overlapping area. The weighted mean trajectory and the weighted variance trajectory are calculated in each window to generate a dynamic trajectory set. In the dynamic trajectory set, time decay reweighting is performed on each trajectory, and samples closer to the current time are assigned higher weights. The trajectory set after calculating the decay factor is obtained as a time-weighted trajectory set. The time-weighted trajectory set is input into the subset selection engine, and samples are filtered based on the sample diversity index and distribution coverage index. When the similarity of samples of the same type in adjacent windows is higher than a set threshold, redundant entries are removed, and only samples that can represent global fluctuation patterns and local abnormal patterns are retained to obtain a dynamic training subset.

[0077] First, each sample in the trusted execution sample library needs to be uniquely indexed by timestamp and device identifier to ensure accurate location later. For each indexed sample, the power execution deviation, energy consumption accumulation deviation, and environmental response deviation are calculated within its corresponding time slice. The power execution deviation is calculated with reference to the power setpoint of the time slice. The actual power within the same time slice is subtracted from the setpoint point by point, the absolute difference is taken, and the average is calculated within the time slice. This average difference is then divided by the average level of the setpoint to obtain a percentage deviation value. To avoid the influence of extreme noise, a very small number of sampling points at the very beginning and end of the time slice can be removed before performing the above steps. The energy consumption accumulation deviation is calculated by accumulating the actual power segment by segment at sampling intervals within the same time slice to obtain the actual energy. Then, the power setpoint is accumulated in the same way to obtain the expected energy. The ratio of the absolute value of the difference between the two to the expected energy represents the energy consumption accumulation deviation of the time slice. Environmental response deviation is used to measure the impact of environmental disturbances on the consistency of execution. Specifically, within the given time slice, the actual trajectories of temperature, humidity, and airflow velocity are compared with reference trajectories from historical stable periods, and point-by-point differences are made. Weights are assigned according to the importance of each factor's impact on energy consumption (e.g., temperature is usually higher than humidity, and humidity is higher than airflow). The time-averaged results of the three types of differences are multiplied by their respective weights, summed, and then normalized as a percentage to obtain the environmental response deviation. These three deviation values ​​are written into a sample index table along with timestamps and device identifiers, serving as a quality scoring vector; smaller values ​​indicate higher quality.

[0078] After obtaining the sample index table with quality scores, it is necessary to retrieve matching baseline samples from the historical training set. Matching is based on equipment category, time slice location, and operating condition similarity: the equipment category must be consistent; the time slice prioritizes the same time period within the same day, and if missing, it is postponed to the same time period of the preceding and following days; operating condition similarity is determined by comparing whether the average load level and temperature level within the reference window fall within a preset error band, and both are considered to be within the error band for a successful match. Each reliable execution sample is paired and fused with its matched baseline sample. During fusion, the quality score vector is used as the weight source, and the three deviation components are mapped to three weights respectively: when a certain deviation is small, a higher weight is assigned to that component, and when the deviation is large, a lower weight is assigned; to avoid a single component dominating, the three weights are proportionally compressed before use, so that the highest weight does not exceed a certain proportion of the sum of all weights. Subsequently, a weighted synthesis is performed on the paired samples field by field: for example, on the power trajectory, the trajectory of the trusted execution sample and the trajectory of the baseline sample are linearly synthesized according to the weights corresponding to the power execution deviation to obtain a reference trajectory that is closer to actual execution and retains the baseline shape; the same synthesis is performed on the energy consumption and environment-related fields according to the weights corresponding to the cumulative energy consumption deviation and the environmental response deviation, respectively. A deviation correction factor is generated during the paired fusion. This factor is obtained by comparing the differences between the trusted execution sample and the baseline sample in core statistics, including the average level, peak position, and duration within the time slice: the amplitude correction factor is obtained by comparing the average level of the trusted execution sample with the average level of the baseline; the phase correction factor is obtained by dividing the difference in the peak occurrence time of the two trajectories by the time slice length; and the persistence correction factor is obtained by dividing the difference in the length of the intervals that continuously exceed the reference threshold by the time slice length. These three factors together constitute the deviation correction factor, which is output as a fused sample sequence along with the fused samples.

[0079] A sliding window is established on the fused sample sequence to obtain a dynamic trajectory set. The window size is set according to the energy consumption fluctuation cycle of the target scene. For example, when the daily cycle is significant, it can be one or two hours, so that at least multiple windows are included in one cycle. The step size is half or less of the window length to form an overlapping area and ensure the continuity of information between adjacent windows. Within each window, the weighted mean trajectory and the weighted variance trajectory need to be calculated. The calculation method of the weighted mean trajectory is as follows: for all fused samples constituting the window, perform the calculation on each time scale of the window time axis, multiply the value of all samples at that scale by the corresponding quality weight and add them up, and then divide by the sum of these quality weights to obtain the mean of that scale; repeat the above steps along each scale of the window to form a complete weighted mean trajectory. The weighted variance trajectory is calculated as follows: On the same scale, subtract the weighted mean of that scale obtained in the previous step from the value of each sample, take the absolute order of magnitude of the difference, multiply it by the same quality weight, and sum them. Finally, divide by the sum of the quality weights to obtain the dispersion of that scale. Connect these values ​​along the time axis to form the weighted variance trajectory. Package the two trajectories of each window with the corresponding bias correction factor to form a dynamic trajectory set.

[0080] For each trajectory in the dynamic trajectory set, a time-decay reweighting process is performed to reflect the "closer is more important" update principle. First, a decay step size is determined, such as in hours or in windows. The weight of the window closest to the current time is used as the starting point, and the weights of earlier windows are gradually reduced by a fixed proportion. This proportion can be set empirically to be slightly below one; for example, the weight is multiplied by a constant slightly less than one for each forward window, until the beginning of the window coverage period. For samples spanning different operating conditions, the operating condition label (e.g., weekdays vs. non-weekdays, peak / valley times) can be checked before decay. If the label is different, the weight is reduced again to avoid the different operating conditions having an excessive impact on the current update. The decayed weights are then renormalized to obtain the trajectory set after decay factoring, i.e., the time-weighted trajectory set.

[0081] The time-weighted trajectory set is input into the subset selection engine, and a dynamic training subset is selected based on both sample diversity and distribution coverage metrics. The sample diversity metric ensures sufficient difference between selected samples. Specifically, it is calculated by comparing the numerical differences of two candidate trajectories on the same time axis, taking the absolute difference, and averaging it across the entire window to obtain a distance metric between the two trajectories. For each candidate trajectory, the distance is calculated pairwise, prioritizing the trajectory whose minimum distance to all trajectories in the selected set is still relatively large. These trajectories are added gradually according to the principle of "farthest to nearest," until a predetermined subset size is reached or the diversity benefit is no longer significant. The distribution coverage metric ensures that key intervals are covered. Specifically, at each time scale, the trajectory values ​​are divided into several intervals from smallest to largest. The position of each candidate trajectory within each interval is recorded as a fingerprint. The coverage number of the selected set across each interval at each time scale is counted. If some intervals are not yet covered, trajectories that can fill these gaps are selected from the candidate set and added first. To avoid redundancy, redundant entries are removed when the similarity of similar samples within adjacent windows exceeds a set threshold. Similarity calculation is the reverse of distance calculation. First, distance is calculated, then mapped to a similarity score between zero and one; the smaller the distance, the greater the similarity. The threshold can be automatically determined using historical data: within a representative historical interval, the similarity distribution of all adjacent windows is calculated, and the score corresponding to a fixed percentile at the top of this distribution is taken as the threshold. Entries with similarities higher than this threshold are considered redundant during actual screening. By prioritizing diversity, then coverage, and finally removing redundancy, a sample set representing both global fluctuation patterns and local anomaly patterns is obtained, forming a dynamic training subset.

[0082] A second embodiment of this application provides an electronic device, the electronic device comprising: processor; The memory is used to store a program, which, when read and executed by the processor, executes the energy consumption prediction and scheduling control method based on machine learning provided in the first embodiment of this application.

[0083] The third embodiment of this application provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it executes a machine learning-based energy consumption prediction and scheduling control method provided in the first embodiment of this application.

[0084] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

Claims

1. A machine learning-based energy consumption prediction and scheduling control method, characterized in that, include: The system collects operating parameters, energy consumption curves, and environmental disturbance data from multiple devices through distributed sensing terminals. The operating parameters, energy consumption curves, and environmental disturbance data are input into a pre-trained machine learning model, which combines a temporal recursive structure and attention weights to generate a probability prediction result of energy consumption distribution. An intervention signal is applied within the equipment safety boundary, and equipment response characteristics are generated based on the intervention differences. These characteristics are then combined with probability prediction results to form an energy consumption risk map. Based on the energy consumption risk map, extreme disturbance scenarios are generated using digital twins. The consistency between the probability prediction results and the scheduling scheme is verified in the data domain and physical domain. Multi-level scheduling is performed based on the verified energy consumption risk map and the task delayability to form the scheduling result. The scheduling results are sent to the devices and feedback data is transmitted back. The machine learning model is then updated based on the feedback data.

2. The energy consumption prediction and scheduling control method based on machine learning according to claim 1, characterized in that, The process of collecting operating parameters, energy consumption curves, and environmental disturbance data from multiple devices via distributed sensing terminals includes: In the distributed sensing terminal, the voltage, current and power signals of the device are sampled at high frequency, and an initial sequence with timestamps is generated during the sampling process; In the distributed sensing terminal, multi-scale wavelet decomposition is performed on the initial sequence to extract energy consumption fluctuation components at different time scales, and the energy consumption fluctuation components are reconstructed into an energy consumption multi-resolution feature sequence. In a distributed sensing terminal, the energy consumption multi-resolution feature sequence is coupled with environmental disturbance parameters to construct an environmental labeled energy consumption vector. The environmental labeled energy consumption vector contains operating parameter features and environmental disturbance weights within the same time slice. In the distributed sensing terminal, a dynamic reliability assessment is performed on the environmental labeled energy consumption vector. A reliability factor is calculated by performing a deviation test with the historical distribution, and the reliability factor is appended to the environmental labeled energy consumption vector to form a reliable energy consumption sample. The data aggregation node receives the trusted energy consumption samples and performs time alignment and spatial fusion on the trusted energy consumption samples from different distributed sensing terminals, outputting a high-integrity energy consumption dataset in a unified format for subsequent machine learning models to call.

3. The energy consumption prediction and scheduling control method based on machine learning according to claim 1, characterized in that, The machine learning model includes a temporal decomposition unit, a two-stream recursive unit, a mask attention unit, a constraint-based interlocking unit, and a probability flow generation unit. The time series decomposition unit is used to receive operating parameters, energy consumption curves and environmental disturbance data, perform timestamp alignment and multi-scale segmentation on the received data, and obtain an initial interval feature sequence. The time series decomposition unit is implemented by wavelet decomposition and sliding window method, so that energy consumption fluctuations and long-term trends at different time scales can be preserved in the initial interval feature sequence at the same time. The dual-stream recursive unit is used to receive the initial interval feature sequence, generate steady-state hidden states and perturbation hidden states respectively, and obtain a joint state sequence through affine coupling. The dual-stream recursive unit sets up a steady-state recursive channel to capture the steady energy consumption evolution law and a perturbation recursive channel to capture the sudden perturbation response. The two channels are synthesized into the joint state sequence at each time step through a nonlinear fusion layer. The mask attention unit is used to receive the joint state sequence and the initial interval feature sequence, generate a context vector, and filter the attention weights by constructing a physical feasibility mask to shield features that do not meet the device's rated power or safety boundary, so that only time segment information that meets the physical constraints is retained after attention normalization, and the context vector is output. The constraint-embedded unit is used to receive the joint state sequence and the context vector, and combine the upper and lower limits of power, energy conservation constraints and climb rate constraints to generate condition codes. The constraint-embedded unit is implemented through condition vector construction and nonlinear projection, so that the condition codes can simultaneously embed statistical prediction laws and physical constraints, ensuring the consistency of subsequent generation results. The probability flow generation unit is used to receive the conditional coding, perform a layer-by-layer reversible transformation on the basic noise distribution and calculate the Jacobian determinant at each layer to obtain the probability prediction result of the energy consumption distribution. The probability flow generation unit is implemented through a reversible probability flow network and provides quantile prediction, confidence interval prediction and probability density function estimation results at the output end. The probability prediction results are used for subsequent energy consumption risk map construction and multi-level scheduling execution.

4. The energy consumption prediction and scheduling control method based on machine learning according to claim 1, characterized in that, The process of applying an intervention signal within the equipment safety boundary, generating equipment response characteristics based on the intervention differences, and combining these characteristics with probability prediction results to form an energy consumption risk map includes: Based on the equipment's rated power, maximum allowable current, temperature rise threshold, and ramp-up rate constraints, the equipment safety boundary is calculated, and the upper limit of the amplitude, duration, rise time, and spectrum limit of the intervention signal are determined accordingly. An intervention signal configuration table is then output. Based on the intervention signal configuration table, the current operating parameters, energy consumption curves, environmental disturbance data, and time synchronization benchmark, an unrelated intervention plan covering different time slices and different devices is generated. The unrelated intervention plan uses mutually orthogonal pseudo-random binary sequences and frequency sweep signals to avoid crosstalk, thereby obtaining the intervention plan and trigger timestamp set. An intervention signal is injected into the device control interface according to the set of trigger timestamps, and operating parameters, energy consumption curves and environmental disturbance data are collected synchronously within a fixed observation window before and after the intervention. Time alignment, detrending and bandpass filtering are performed to obtain the intervention window dataset. The intervention window dataset is used to estimate the response. The impulse response, gain, delay, time constant, recovery time and steady-state offset are calculated by a joint method of subspace identification and deconvolution. The confounding effects of environmental disturbance data on the response estimation are eliminated by instrumental variable regression to obtain the equipment response characteristics. The device response features are fused with the probability prediction results, and the distribution parameters of the probability prediction results are corrected based on likelihood weighting and Bayesian update to obtain the corrected probability prediction results. Using equipment and time slices as indexes, the out-of-limit probability, variance, and tail thickness of the corrected probability prediction results are compared with the equipment safety boundary to generate map entries containing risk intensity, source tags, and time location, and aggregated according to the index to form an energy consumption risk map.

5. The energy consumption prediction and scheduling control method based on machine learning according to claim 1, characterized in that, The process involves generating extreme disturbance scenarios using digital twins based on energy consumption risk maps, and verifying the consistency between probability prediction results and scheduling schemes in both the data and physical domains. Multi-level scheduling is then performed based on the verified energy consumption risk maps and task delayability to form the scheduling result, including: Based on the energy consumption risk map, high-risk areas and potential abnormal areas are extracted and marked. A risk triggering condition set is generated by combining operating parameters, energy consumption curves and environmental disturbance data. A digital twin simulation environment is constructed based on the risk triggering condition set. The physical characteristics of the equipment, energy consumption curves and environmental disturbance models are loaded into the simulation environment, and the marked areas are initialized to obtain the initial state of the simulation. Extreme disturbance trajectories are injected into the initial state of the simulation. These extreme disturbance trajectories include scenarios such as power surge, temperature change, voltage fluctuation, and concurrent switching of multiple devices, generating a corresponding simulation response dataset. The simulation response dataset and the probability prediction result are compared in the data domain, the mean, variance and quantile deviation are calculated, and the probability prediction result is corrected when the deviation exceeds a set threshold to obtain the probability prediction result corrected by the data domain. The probability prediction results corrected by the data domain are verified against the scheduling scheme in the physical domain to check whether the scheduling scheme meets the constraints of equipment rated power, maximum allowable current, energy conservation and ramp rate, and to obtain the scheduling scheme verified by the physical domain. By combining the probability prediction results corrected by the data domain with the scheduling scheme verified by the physical domain, the energy consumption risk map is updated and the task delayability is marked. Multi-level scheduling is then performed according to the task delayability to obtain the confirmed scheduling results.

6. The energy consumption prediction and scheduling control method based on machine learning according to claim 1, characterized in that, The step of sending the scheduling results to the device and transmitting feedback data back, and updating the machine learning model based on the feedback data, includes: The confirmed scheduling results are converted into a set of energy consumption management instructions that can be recognized by the device. The set of energy consumption management instructions includes power settings, task execution time windows, load allocation ratios and safety redundancy parameters. It is then sent to the target device through an encrypted communication protocol with digital signatures and timestamps to obtain a record of the sent instructions with a unique identifier. On the device side, energy consumption control operations are performed according to the issued instructions. During the execution process, operating parameters, energy consumption curves and environmental disturbance data are collected in real time to generate an execution feedback stream with an instruction identifier. The execution feedback stream is transmitted back to the data aggregation node through the same encrypted communication protocol to obtain a feedback dataset bound by the identifier. At the data aggregation node, the feedback dataset and the issued instruction record are timestamped and their integrity is verified. If the verification result shows that the feedback dataset and the issued instruction record are consistent, the feedback dataset is stored as a trusted execution sample. During the model update phase, the trusted execution sample library is fused with the historical training set of the machine learning model, and the execution data of the most recent time period is selected using the sliding window method to construct a dynamic training subset. Incremental training is performed on the dynamic training subset in the machine learning model. The model parameters are corrected by minimizing the probability distribution difference and Bayesian regularization to generate an online updated machine learning model. The online updated machine learning model is then used for the probability prediction of subsequent energy consumption distribution to obtain a closed-loop adaptive prediction model.

7. The energy consumption prediction and scheduling control method based on machine learning according to claim 6, characterized in that, The process of fusing the trusted execution sample library with the historical training set of the machine learning model during the model update phase, and using a sliding window method to select the execution data from the most recent time period to construct a dynamic training subset, includes: Each sample in the trusted execution sample library is indexed based on timestamp and device identifier, and power execution deviation, energy consumption cumulative deviation and environmental response deviation are calculated within the same time slice. The three are combined into a quality score vector to obtain a sample index table with quality scores. According to the sample index table, the baseline samples corresponding to the time slice and device category are retrieved in the historical training set, and the trusted execution samples are paired and fused with the baseline samples using the quality score vector as the weight to generate a fused sample sequence with a bias correction factor. A sliding window is constructed on the fused sample sequence. The window size is set according to the energy consumption fluctuation cycle of the target scene. The window step size is smaller than the window size to form an overlapping area. The weighted mean trajectory and the weighted variance trajectory are calculated in each window to generate a dynamic trajectory set. In the dynamic trajectory set, time decay reweighting is performed on each trajectory, and samples closer to the current time are assigned higher weights. The trajectory set after calculating the decay factor is obtained as a time-weighted trajectory set. The time-weighted trajectory set is input into the subset selection engine, and samples are filtered based on the sample diversity index and distribution coverage index. When the similarity of samples of the same type in adjacent windows is higher than a set threshold, redundant entries are removed, and only samples that can represent global fluctuation patterns and local abnormal patterns are retained to obtain a dynamic training subset.

Images