Multi-energy load forecasting and low-carbon optimization regulation method

Abstract

The application discloses a multi-energy load prediction and low-carbon optimization regulation method, collects multi-source terminal load sequences, and obtains unified load sequence samples through repair and alignment; the unified load sequence samples are subjected to multi-frequency decomposition, and four types of components, i.e., trend, period, pulse and high-frequency fluctuation, are obtained; the four types of components, exogenous characteristics, time position coding, load type coding are input into a multi-energy load interval prediction model to predict a median prediction sequence and a multi-energy load prediction interval; the median prediction sequence is used as a terminal demand benchmark, external supply energy benchmark consumption across the park boundary is determined according to the balance of electricity, cold, heat, gas and hydrogen energy, and carbon emissions are accounted, and a comprehensive carbon emission trend is formed by summarizing; low-carbon suggestions are generated according to the prediction interval and the carbon emission trend, and are sent to an external dispatching control system for execution. The method can improve the alignment of multi-source asynchronous data, the expression of multi-frequency characteristics, the accuracy of multi-energy coupling prediction and carbon accounting, and provide reliable basis for low-carbon optimization regulation.

Description

Technical Field

[0001] This invention relates to the fields of integrated energy system time series analysis, machine learning prediction, and low-carbon regulation technology, specifically to a method for multi-energy load prediction and low-carbon optimization regulation. Background Technology

[0002] Integrated energy systems typically include electricity load, cooling load, heating load, gas consumption, hydrogen energy devices, energy storage devices, and various energy conversion equipment. Due to differences in sampling periods, reporting delays, data integrity, and dimensions among different data sources, directly inputting raw data into the prediction model can easily lead to state mismatches within the same time period and reduce the reliability of load prediction results. Multi-energy load sequences simultaneously exhibit slow changing trends, stable periods such as daily or weekly cycles, pulses caused by equipment start-ups / shutdowns or sudden events, and continuous high-frequency fluctuations. Existing methods that only use a single time-domain model or only perform ordinary frequency-domain transformations struggle to simultaneously distinguish these time-series components. Furthermore, electricity, cooling, heating, gas, hydrogen, and energy storage are dynamically correlated through heat pumps, chillers, boilers, electrolyzers, and energy storage charging and discharging; single-energy prediction methods are insufficient to express the coupling relationships between multiple energy sources.

[0003] Furthermore, conventional load forecasting results cannot directly reflect low-carbon regulation needs. The park's main electricity and gas meters may already include the input energy from chillers, heat pumps, boilers, electrolyzers, and energy storage devices; repeatedly calculating these device inputs based on cooling and heating loads and conversion efficiencies easily leads to redundant measurement of energy consumption and carbon emissions. Therefore, a clear statistical boundary needs to be established between the end-user demand side, the energy conversion equipment side, and the energy supply side outside the park boundary. Existing low-carbon regulation schemes tend to directly mix forecasting, carbon accounting, and equipment optimization into a complex scheduling model, resulting in unclear and difficult-to-transfer forecasting innovations. This invention takes multi-energy load range forecasting as its core, treats carbon trend calculation and low-carbon suggestion generation as independent subsequent steps, and limits the external scheduling control system to the suggestion execution framework, thereby forming a clear data flow and module interface. Summary of the Invention

[0004] To address the problems in existing technologies, such as the difficulty in unifying and aligning multi-source asynchronous data, the difficulty in accurately separating multi-frequency time-series components, the difficulty in expressing dynamic correlations among multiple energy sources, the difficulty in converting load forecasting results into energy-specific carbon trends, and the lack of unified triggering and generation criteria for low-carbon recommendations, this invention proposes a multi-energy load forecasting and low-carbon optimization control method. The specific technical solution is as follows:

[0005] A method for multi-energy load forecasting and low-carbon optimization control includes the following steps:

[0006] S1 acquires terminal multi-energy load, meteorological and time information, equipment operating status and time-varying data, as well as park boundary and carbon accounting parameters;

[0007] S2 performs anomaly identification, missing data filling, timestamp correction, and unified time grid alignment on the data obtained from S1 to obtain a unified load sequence sample composed of terminal electricity, cooling, heating loads, and direct gas loads, exogenous features, and external energy supply carbon emission factors and time-of-use energy prices aligned by time period;

[0008] S3, perform multi-frequency decomposition on the unified load sequence sample to obtain four types of components: trend component, periodic component, pulse component and high-frequency fluctuation component.

[0009] S4. Input the four types of components, exogenous features, time location codes and load type codes into the multi-energy load interval prediction model. After parallel extraction of embedding splicing, time domain dependent features, adaptive frequency domain features and multi-energy relationship features and adaptive residual fusion, the median prediction sequence and multi-energy load prediction interval are obtained.

[0010] S5. Using the median prediction sequence as the terminal demand benchmark, combined with the park boundary and carbon accounting parameters, and according to the balance of electricity, cooling, heating, gas and hydrogen energy, the external energy supply benchmark consumption that crosses the park boundary is determined and multiplied by the corresponding carbon emission factor to obtain the carbon emission contribution. Then, the carbon emission trend is summarized by time period.

[0011] S6. Based on the multi-energy load forecast range and the comprehensive carbon emission trend, generate low-carbon recommendations and send them to the external dispatch control system for execution.

[0012] Furthermore, anomaly identification involves identifying and marking data with out-of-limit values, long-term constant values, repeated timestamps, and anomalous mutations as invalid observations.

[0013] Missing data filling involves filling in the original missing data and data gaps formed by invalid observations. Short-term gaps are filled using linear interpolation, while long-term gaps are filled using historical data from similar periods of the same energy type.

[0014] Timestamp correction involves adjusting the timestamp based on the sampling period, reporting delay, and device clock offset.

[0015] Unified time grid alignment is as follows: a unified time grid is constructed with a preset time granularity, and the corrected data is aggregated, interpolated, or resampled to the unified time grid to obtain a unified load sequence sample composed of terminal electrical load, terminal cooling load, terminal heating load, and direct gas load.

[0016] Furthermore, S3 specifically includes: independently performing trend extraction, period identification and reconstruction, pulse identification, and high-frequency margin separation on the terminal electrical load, terminal cooling load, terminal heating load, and direct gas load sequences, respectively;

[0017] The trend extraction method is as follows: trend components are extracted by averaging through a sliding window.

[0018] The periodic identification and reconstruction are as follows: performing fast Fourier transform, power spectral density calculation, Top-K positive and negative frequency conjugate peak pair screening, and fast Fourier inverse transform on the residual components after removing the trend components to reconstruct the periodic components;

[0019] The pulse identification and high-frequency residual separation are performed as follows: robust threshold identification based on the median absolute deviation is performed on the remaining residual after removing the periodic component, the pulse component is extracted and removed, and continuous high-frequency fluctuation components are obtained.

[0020] Furthermore, the multi-energy load interval prediction model of S4 includes: an input embedding layer, a time-domain dependent feature extraction unit, an adaptive frequency-domain feature extraction unit, a multi-energy relationship feature extraction unit, a cross-branch adaptive residual fusion unit, and an ordered quantile output layer;

[0021] The input embedding layer maps the trend component, periodic component, pulse component and high-frequency fluctuation component to a unified feature dimension, and then concatenates the mapped exogenous features with the mapping results of the four types of load components, time position code and load type code along the channel dimension to obtain the embedding features with terminal electrical load, terminal cooling load, terminal heating load and direct gas load as the four load nodes.

[0022] The embedded features are processed in parallel into three branches: a time-domain dependent feature extraction unit, an adaptive frequency-domain feature extraction unit, and a multi-energy relationship feature extraction unit, which respectively obtain time-domain features, frequency-domain features, and cross-energy correlation features, and all three branches maintain the same 4-load node axis;

[0023] The cross-branch adaptive residual fusion unit calculates the weights of three branches based on time-domain features, frequency-domain features, and cross-energy correlation features on a unified 4-load node axis. It then performs fusion mapping on the weighted branch features and adds them to the input embedded feature residuals. Finally, it obtains the fused prediction features through normalization.

[0024] The ordered quantile output layer inputs fused prediction features, first forming a median prediction result, and then forming low quantile and high quantile prediction results through non-negative offset, such that the low quantile prediction result is smaller than the median prediction result and the median prediction result is smaller than the high quantile prediction result.

[0025] Furthermore, the temporal-dependent feature extraction unit employs dilated causal convolution, so that the output of time period t does not use observations after time period t, and the dilation rate increases with the number of layers to expand the temporal receptive field; the convolution output is added to the embedded feature residual after nonlinear activation to obtain the temporal feature.

[0026] Furthermore, the adaptive frequency domain feature extraction unit includes a frequency domain main path and a pulse retention bypass path;

[0027] In the frequency domain main path, the embedded features are first transformed to the frequency domain by Fast Fourier Transform (FFT). Then, the low-frequency, mid-frequency, and high-frequency bands are divided by learnable frequency band boundaries or learnable masks, and trend enhancement, period enhancement, and useful high-frequency fluctuation selection are performed respectively. After that, the inverse FFT of each frequency band is performed. In the inverse FFT, conjugate symmetric constraints are applied to the frequency weights so that the low-frequency, mid-frequency, and high-frequency features obtained by the inverse transform are all real-valued time-domain features.

[0028] The pulse retention bypass directly reads the channel formed by independent mapping of pulse components, and retains abrupt change information through pulse gating to obtain pulse characteristics;

[0029] The time-domain features of the low-frequency, mid-frequency, and high-frequency bands after inverse transformation are spliced ​​and adaptively fused with the pulse features, and then added to the residual of the original embedded features to obtain the frequency domain features.

[0030] Furthermore, the multi-energy relationship feature extraction unit constructs a multi-energy relationship graph based on the embedding features of the four load nodes corresponding to the terminal electrical load, terminal cooling load, terminal heating load, and direct gas load, as well as the physical connection relationships of the equipment. The physical connection mask only allows actual electric-cooling, electric-heating, gas-heating, and gas-electric conversion paths between the four types of load nodes. The diagonal elements of the mask and the allowed node pairs are set to zero, and the remaining off-diagonal elements are set to negative infinity. Between the allowed node pairs, a relationship bias is generated based on historical correlation, the start / stop status of energy conversion equipment, load rate or availability status, and hydrogen energy and energy storage status. The embedding features of the nodes are mapped to queries, keys, and values. The relationship weights between nodes are calculated, and the physical connection mask and relationship bias are superimposed before normalization to obtain the final relationship weight matrix. Then, graph attention aggregation is performed to output the cross-energy association features corresponding to the four types of loads.

[0031] Furthermore, S5 specifically includes: taking the median forecast results of terminal electrical load, terminal cooling load, terminal heating load and direct gas load as the terminal demand benchmark;

[0032] Five types of energy balance are established under the baseline operation plan: power balance, cooling balance, heating balance, gas balance and hydrogen energy balance; the baseline operation plan is the currently confirmed operation plan. When there is no currently confirmed operation plan, it is composed of preset energy supply paths and corresponding equipment output, energy storage status and external energy consumption.

[0033] In the aforementioned power balance, grid-purchased electricity, local photovoltaic power, gas-fired combined heat and power generation, and energy storage and discharge are considered as supply items, while terminal electrical load, chillers, heat pumps, electrolyzers, and energy storage and charging are considered as demand items.

[0034] In the aforementioned cooling balance, the cooling supply from the chiller and the cooling release from the cold storage meet the terminal cooling load and the cooling charge from the cold storage.

[0035] In the heat balance, gas boilers, heat pumps, gas combined heat and power, and thermal storage heat release meet the terminal heat load and thermal storage heat charging.

[0036] In the aforementioned gas balance, the purchased gas meets the gas input requirements for direct gas load, gas boilers, and gas-fired combined heat and power.

[0037] In the aforementioned hydrogen energy balance, purchased hydrogen, electrolytic hydrogen production, and hydrogen storage and release meet the requirements for hydrogen energy usage and hydrogen storage and charging.

[0038] The baseline consumption of electricity purchased from the grid, gas purchased from outside, and hydrogen purchased from outside the park is directly obtained from the above five categories of energy balance. For each type of external energy supply that crosses the park boundary, the baseline consumption is multiplied by the corresponding carbon emission factor to obtain the carbon emission contribution of the energy supply, and the results are summarized by time period to form a comprehensive carbon emission trend.

[0039] Furthermore, the low-carbon recommendation generation in S6 specifically includes:

[0040] Based on a calibration dataset, the urgency of suggestions and the confidence of predictions are obtained through normalization with fixed upper and lower bounds and out-of-bounds truncation. The urgency threshold is determined according to the preset high urgency trigger ratio, and the confidence threshold is determined according to the allowed low confidence ratio, and is fixed within the current application cycle.

[0041] Upper-level low-carbon economy optimization is triggered only when the urgency of the recommendation reaches the urgency threshold and the confidence of the prediction reaches the confidence threshold.

[0042] The upper-level low-carbon economic optimization takes the transferable load adjustment, energy storage charging and discharging power, energy conversion equipment output and flexible equipment start-up and shutdown status as decision variables, and takes the dimensionless operating cost relative to the benchmark operation plan and the coordinated reduction of carbon emissions from the external energy supply side as optimization objectives, while satisfying five categories of energy balance, equipment output and ramp-up, energy storage status, transferable load conservation and allowable time period constraints.

[0043] The optimization results are analyzed and combined with the comprehensive carbon trend to generate a set of regulation suggestions and carbon signal push suggestions. The set of regulation suggestions includes load transfer suggestions, energy storage charging and discharging suggestions, energy conversion suggestions, and equipment operation adjustment suggestions. The carbon signal push suggestions are generated based on the high carbon period, low carbon period, and urgency of the suggestions in the comprehensive carbon emission trend.

[0044] Furthermore, for each regulation suggestion in the set of regulation suggestions, a single suggestion counterfactual recalculation is performed to determine the marginal contribution based on the difference in carbon emissions and costs between the counterfactual scheduling result without the suggestion and the optimal scheduling result. The original score of the regulation suggestion is calculated based on the expected carbon emission reduction, time urgency, implementation feasibility, and positive cost increment. The original score of the carbon signal push suggestion is calculated based on the time urgency and information push feasibility. The two types of original scores are truncated and normalized to a uniform scale and arranged in descending order to obtain the priority of the suggestions.

[0045] The present invention has the following beneficial effects:

[0046] (1) By measuring the boundary of the park and splitting the energy flow, we can distinguish the terminal demand, the input of the energy conversion equipment and the external energy supply. By identifying anomalies, filling missing information, correcting timestamps and aligning the unified time grid, we can form a unified load sequence sample with the same time semantics, which can provide a consistent and reliable data foundation for subsequent load decomposition, forecasting, carbon accounting and low-carbon recommendations.

[0047] (2) By peeling away the four components of trend, period, pulse and high-frequency fluctuation step by step, it is possible to distinguish slow changes, stable periods, isolated mutations and continuous rapid fluctuations, thereby improving the interpretability of multi-frequency time series characteristics.

[0048] (3) By parallel extraction of time-domain dependent features, adaptive frequency domain features and cross-energy correlation features and adaptive residual fusion, it can simultaneously express time dependence, key frequency bands and multi-energy dynamic correlation, thus completing multi-energy load forecasting more accurately, more robustly and more interpretably.

[0049] (4) By outputting ordered quantiles, low, medium and high quantile prediction results are generated. While providing the central prediction value and prediction interval, quantile crossover is avoided, thus preventing errors in subsequent carbon trend calculation, urgency judgment and low carbon optimization triggering.

[0050] (5) Determine the benchmark usage of external energy supply across the park boundary through a clear balance of electricity, cooling, heating, gas and hydrogen, and perform carbon accounting only on the side of external energy supply to avoid duplicate carbon accounting for terminal cooling, heating, energy storage and self-produced hydrogen with purchased electricity or purchased gas.

[0051] (6) Through index normalization with fixed boundaries and over-boundary cutoff rules, dual thresholds for recommendation urgency and prediction credibility, and low-carbon economic optimization with clear supply and demand, equipment, energy storage and load transfer constraints, five types of structured recommendations that can be ranked and executed are generated. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the overall process of a multi-energy load prediction and low-carbon optimization control method according to an embodiment of the present invention;

[0053] Figure 2 This is a flowchart of the multi-frequency decomposition module processing according to an embodiment of the present invention;

[0054] Figure 3 This is a flowchart illustrating the structure and processing flow of the multi-energy load range prediction module according to an embodiment of the present invention.

[0055] Figure 4 This is a flowchart of the adaptive frequency domain feature extraction unit in an embodiment of the present invention;

[0056] Figure 5 This is a flowchart of the carbon trend calculation module in an embodiment of the present invention;

[0057] Figure 6 This is a flowchart of the low-carbon suggestion module in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0059] like Figure 1 As shown, the present invention proposes a multi-energy load forecasting and low-carbon optimization control method. The overall process includes: multi-source data acquisition and multi-frequency characterization, multi-energy load interval forecasting and carbon trend calculation, low-carbon suggestion generation, and external scheduling control. Specifically, it includes the following steps:

[0060] Step S1: Obtain terminal multi-energy load, meteorological and time information, operating status and time-varying data, as well as park boundary and carbon accounting parameters; the terminal multi-energy load includes terminal electrical load, terminal cooling load, terminal heating load and direct gas load; the operating status and time-varying data include equipment and its input and output information, hydrogen energy and energy storage status, and energy price;

[0061] Step S2 involves anomaly identification, missing data filling, timestamp correction, and unified time grid alignment of the data obtained in step S1 to obtain unified load sequence samples, exogenous features, and external energy supply carbon emission factors and time-of-use energy prices aligned by time period.

[0062] Step S3: Perform multi-frequency decomposition on the uniform load sequence sample to obtain four components: trend, period, impulse, and high-frequency fluctuation.

[0063] Step S4: Combine the above four components with exogenous features such as weather, time, equipment, hydrogen energy and energy storage status, and use the multi-energy load interval prediction model to predict the median prediction sequence and the multi-energy load prediction interval.

[0064] Step S5: Using the median prediction sequence as the terminal demand benchmark, and combining the park boundary, external energy supply carbon emission factor, hydrogen energy and energy storage status and energy changes, solve the sub-energy balance according to electricity, cooling, heating, gas and hydrogen energy, thereby determining the external energy supply benchmark usage that crosses the park boundary in each time period and multiplying it by the corresponding carbon emission factor to obtain the carbon emission contribution, and summarizing by time period to form a comprehensive carbon trend.

[0065] Step S6: Generate low-carbon recommendations based on the multi-energy load forecast range and the comprehensive carbon emission trend, and send the low-carbon recommendations to the external dispatch control system for execution.

[0066] In one embodiment, a park integrated energy system is taken as the object, which includes photovoltaic power generation devices, gas turbines, heat pumps, chillers, boilers, energy storage devices, thermal storage devices, hydrogen energy devices, and user-side flexible loads. The system can use fifteen minutes as a uniform time granularity and the next twenty-four time periods as the prediction time domain; the time granularity and prediction time domain are example parameters and do not constitute a limitation of the present invention.

[0067] Based on the above-mentioned multi-energy load forecasting and low-carbon optimization control methods, this invention designs a corresponding multi-energy load forecasting and low-carbon optimization control system. Each method step is implemented using the corresponding modules in the system, including: a data acquisition module, a time alignment module, a multi-frequency decomposition module, a multi-energy load interval forecasting model, a carbon trend calculation module, and a low-carbon suggestion module.

[0068] The data acquisition module uses the physical metering boundaries between the park and the public power grid, external gas network, and external hydrogen source as the system boundary. It acquires data uploaded from terminal energy consumption sub-meters, energy conversion equipment input / output sub-meters, park boundary energy gate metering devices, hydrogen energy metering devices, energy storage monitoring devices, weather stations, equipment controllers, carbon emission factor interfaces, and energy price interfaces. Specifically, this includes: historical terminal electrical load, terminal cooling load, terminal heating load, direct gas load, temperature, humidity, solar irradiance, calendar characteristics, equipment and its operating status, energy storage state of charge, charging and discharging power, hydrogen storage or hydrogen flow, electricity purchased from the park boundary power grid, purchased gas and hydrogen, time-varying external energy supply carbon emission factors, and time-of-use energy prices.

[0069] The above data is then used to form dynamic time-series data, specifically by recording the sampling timestamp, reporting timestamp, data quality markers, and equipment status of each dynamic data source. The time-varying external energy carbon emission factors and time-of-use energy prices are then incorporated into the multi-source asynchronous time-series data according to their corresponding timestamps to form a terminal load sequence.

[0070] The park boundary definition, physical connection relationship of equipment, equipment conversion efficiency, baseline operation plan, equipment start-up and shutdown cost, operation and maintenance cost, and executable constraints are used as configuration parameters that are statically or periodically updated. No missing data filling or unified time grid resampling is performed on these configuration parameters; updates are only performed when the parameters take effect. The baseline operation plan is a currently confirmed operation plan. When no currently confirmed operation plan exists, it consists of a preset power supply path and the corresponding equipment output, energy storage status, and external energy consumption.

[0071] In addition, when the terminal demand lacks an independent sub-table, that is, only the park master table data is available, the master table data is split into energy flow based on the energy conversion equipment input sub-table, equipment operating status and conversion efficiency to form a terminal load sequence.

[0072] The time alignment module first identifies out-of-limit values, long-term constant values, duplicate timestamps, and anomalous mutations in the multi-source asynchronous time-series data. Identified anomalous data is marked as invalid observations, and anomaly validity markers are generated. Next, it fills in the original missing data and data gaps formed by invalid observations. Short-term gaps are filled using linear interpolation, while long-term gaps are filled using historical data from similar periods of the same energy type, and missing data filling markers are generated. Finally, timestamps are corrected based on the sampling period, reporting delay, and device clock deviation. Finally, the corrected data is mapped to a unified time grid to obtain a time series matrix composed of energy types, while retaining the alignment results of data validity, delay, and missing data markers. Specifically, a unified time grid is constructed with a preset time granularity. The corrected dynamic time series data is aggregated, interpolated, or resampled to the unified time grid to obtain a unified load series sample composed of terminal electrical load, terminal cooling load, terminal heating load, and direct gas load. Exogenous features such as meteorological, time, equipment information, hydrogen energy, and energy storage status are obtained, along with external energy supply carbon emission factors and time-of-use energy prices aligned by time period, while retaining data validity markers, delay markers, and missing data markers. The unified time grid can be represented as:

[0073] ,

[0074] ,

[0075] ,

[0076] in, Indicates the first A unified point in time; Indicates the start time; Indicates the duration of time between adjacent points in time; Indicates the quantity at a uniform point in time; This indicates that load type m is at a uniform point in time. Alignment value at; This represents the i-th original timestamp of load type m; The original set of observation sequences represents the load type m at the original timestamp. Observations at; (·) indicates timestamp correction, aggregation, interpolation, and validity marker generation operations.

[0077] The multi-frequency decomposition module, such as Figure 2 As shown, the algorithm only applies to the aligned uniform load sequence sample, with meteorological, temporal, equipment, hydrogen energy, and energy storage status treated as exogenous features to bypass multi-frequency decomposition. Specifically, for the terminal electrical load, terminal cooling load, terminal heating load, and direct gas load sequences, trend extraction, period identification and reconstruction, pulse identification, and high-frequency margin separation are performed independently.

[0078] For example, the trend component is extracted from the terminal electrical load sequence using a sliding window average. The residual component after removing the trend component is then reconstructed using Fast Fourier Transform (FFT), power spectral density calculation, Top-K positive and negative frequency conjugate peak pair selection, and Inverse Fast Fourier Transform (IFFT) to reconstruct the periodic component. , can be represented as:

[0079] ,

[0080] ,

[0081] ,

[0082] ,

[0083] in, Represents the detrended residual components The complex spectrum, Represents the power spectral density. This indicates the number of sample points within the decomposition window. This indicates the number of significant spectral peak pairs retained. This represents the set of frequencies that are conjugate closed with respect to positive and negative frequencies; This indicates the operation of selecting K sets of conjugate spectral peak pairs according to the power spectral density. The conjugate coefficients are retained in pairs to ensure that the periodic reconstruction result is a real-valued sequence; f represents the frequency variable.

[0084] Robust threshold identification based on median absolute deviation (MAD) is performed on the remaining residual after removing the periodic component to extract the pulse component. Then, these components are removed to obtain continuous high-frequency fluctuation components. , can be represented as:

[0085] ,

[0086] ,

[0087] ,

[0088] in, This represents the residual after removing the trend and periodic components. This represents the median of the remaining residuals. The median absolute deviation of the remaining residuals is represented by κ, which represents the robustness threshold coefficient. The pulse component is used to characterize isolated spikes or short-term abrupt changes, while the high-frequency fluctuation component is used to characterize continuous and rapid changes after pulse removal.

[0089] The window length, number of peak pairs, and threshold coefficient are determined using a training set or a validation set.

[0090] The multi-frequency decomposition module completes the four-component multi-frequency decomposition using a step-by-step stripping method, which can be represented as:

[0091] ,

[0092] ,

[0093] in, This represents a unified time-series sample of load type m in time period t, where m only includes terminal electrical load, terminal cooling load, terminal heating load, and direct gas load types. Indicates trend components, This represents the r-th periodic component. This represents the sum of the periodic components. Indicates the pulse component. This represents high-frequency fluctuation components.

[0094] The multi-energy load range prediction model, such as Figure 3 As shown, it includes: an input embedding layer, a time-domain dependent feature extraction unit, an adaptive frequency domain feature extraction unit, a multi-energy relationship feature extraction unit, a cross-branch adaptive residual fusion unit, and an ordered quantile output layer.

[0095] The input embedding layer maps four types of load components—trend component, periodic component, pulse component, and high-frequency fluctuation component—to a unified feature dimension. It then concatenates the exogenous feature mappings of weather, time, equipment, hydrogen energy, and energy storage status with the mapping results of the four load components, time location encoding, and load type encoding along the channel dimension to obtain the embedding features with four load nodes: terminal electrical load, terminal cooling load, terminal heating load, and direct gas load. This can be represented as:

[0096] ,

[0097] ,

[0098] ,

[0099] in, This represents the embedding characteristics of load type m in time period t. , , and These represent the mapping parameters for trend, period, impulse, and high-frequency fluctuation components, respectively. Indicates time location encoding. Indicates the load type encoding. This represents the exogenous characteristics of meteorology, equipment, hydrogen energy, and energy storage status during the same period. The parameters represent the exogenous feature mapping. Concat(·) represents concatenation performed along the feature channel dimension, and Stack(·) represents stacking performed along the load node dimension. This represents a set of four types of load nodes consisting of terminal electrical load, terminal cooling load, terminal heating load, and direct gas load. e, c, h, and g represent the terminal electrical load, terminal cooling load, terminal heating load, and direct gas load nodes, respectively. This indicates a multi-energy embedding feature with a node axis length of 4 and retaining four independent component channels; hydrogen energy and energy storage status are only spliced ​​to the corresponding load nodes as exogenous features and do not occupy node axes separately.

[0100] The embedded features are processed in parallel into three branches: a time-domain dependent feature extraction unit, an adaptive frequency-domain feature extraction unit, and a multi-energy relationship feature extraction unit. All three branches maintain the same 4-load node axis.

[0101] The temporal-dependent feature extraction unit employs dilated causal convolution to ensure that the output for time period t does not use observations after time period t. The dilation rate increases with the number of layers to expand the temporal receptive field. The convolutional output, after nonlinear activation, is added to the residual of the input embedded features. This can be expressed as:

[0102] ,

[0103] ,

[0104] in, This represents the output of a dilated causal convolution. and Represents the convolution parameters. σ(·) represents a causal convolution with an inflation rate of d, and σ(·) represents a nonlinear activation function, using the ReLU activation function. Indicates the output mapping parameters. The formula represents the temporal features. In the formula, t represents the time period index, d represents the causal convolution dilation rate, and E with subscript t represents the multi-energy embedding feature of time period t; the independent variable of σ(·) is the linear output of the convolution.

[0105] The time-domain dependent feature extraction unit can use a single simplified residual block or multiple stacked residual blocks.

[0106] The adaptive frequency domain feature extraction unit includes a frequency domain main path and a pulse-preserving bypass path. For example... Figure 4 As shown, in the main frequency domain path, the embedded features are first transformed to the frequency domain using Fast Fourier Transform (FFT). Then, low-frequency, mid-frequency, and high-frequency bands are divided using learnable frequency band boundaries or learnable masks, and trend enhancement, period enhancement, and useful high-frequency fluctuation selection are performed respectively. Finally, inverse transform (IFFT) is performed on each frequency band. This can be represented as:

[0107] ,

[0108] ,

[0109] in, Frequency domain representation of embedded features Represents embedded features, This represents the learnable mask or frequency weight corresponding to frequency band b. Let represent the time-domain features obtained by the inverse fast Fourier transform of frequency band b, where L, M, and H represent low frequency, mid frequency, and high frequency, respectively, and ⊙ denotes element-wise multiplication. To ensure that the low-frequency, mid-frequency, and high-frequency features obtained by the inverse fast Fourier transform are all real-valued sequences, a conjugate symmetric constraint is applied to the frequency weights: , Let f represent the complex conjugate of the corresponding frequency weights, and let f represent the frequency variable. Since the input embedded features are real-valued sequences, their positive frequency coefficients and corresponding negative frequency coefficients are complex conjugates. Therefore, the spectrum after element-wise multiplication of the frequency weights and the frequency domain representation still satisfies the conjugate symmetry relationship. After performing the inverse fast Fourier transform, the real-valued time-domain features are obtained. When the frequency weights are real-valued, the above constraint is equivalent to the weights at positive frequencies being equal to the weights at corresponding negative frequencies.

[0110] Simultaneously, the channel formed by independent mapping of pulse components is directly read via pulse retention bypass. Furthermore, pulse characteristics are obtained by preserving mutation information through pulse gating.

[0111] Finally, the three features obtained from the inverse transformation are concatenated and adaptively fused with the pulse features preserved after bypass mapping, and then added to the residual of the original embedded features to obtain the frequency domain features. This can be represented as:

[0112] ,

[0113] ,

[0114] in, Indicates the pulse gating weight. This represents the channel corresponding to the pulse component in the embedded features. This represents the preserved pulse features, and Concat(·) represents multi-scale feature concatenation. This represents an adaptive fusion mapping. This represents the frequency domain characteristics. In the formula, , , These represent low-frequency, mid-frequency, and high-frequency characteristics, respectively. This indicates that an adaptive fusion mapping will be performed on the splicing result.

[0115] The pulse component extraction based on median absolute deviation, the pulse retention bypass for directly reading pulse mapping channels, and the physical connection mask constructed according to the actual energy conversion paths in the park are all improved structures designed to address the problems of multi-energy load spikes being easily smoothed in the frequency domain and the non-arbitrary connectivity between different energy nodes. These are not inherent modules of the original MFformer or MSGNet models, nor are they direct references to their original formulas. Specifically, pulse component extraction distinguishes between isolated abrupt changes and continuous high-frequency fluctuations before model input; the pulse retention bypass allows abrupt changes to bypass the frequency band smoothing process; and the physical connection mask is used to exclude node relationships in the park that do not have actual energy conversion paths.

[0116] The multi-energy relationship feature extraction unit constructs a multi-energy relationship diagram based on the characteristics of four load nodes corresponding to terminal electrical load, terminal cooling load, terminal heating load, and direct gas load, as well as the physical connection relationships of the equipment. Node features consist of the corresponding load component, equipment status, and availability; hydrogen energy and energy storage status are used as exogenous features to participate in the attribute mapping or relationship biasing of relevant load nodes. The calculation does not constitute a node individually, nor does it form node pairs in the physical connection mask. Off-diagonal node pairs in the physical connection mask are limited to actual existing electricity-cooling, electricity-heating, gas-heating, and gas-electricity conversion paths within the park. Graph attention aggregation is performed after superimposing historical correlations or energy conversion relationship biases between allowed node pairs. The aggregation process satisfies:

[0117] ,

[0118] ,

[0119] ,

[0120] ,

[0121] ,

[0122] in, , and These represent the query, key, and value feature matrices, respectively, with the superscript T indicating matrix transpose. , and Indicates the corresponding mapping parameters. The set of four load nodes represents the weight matrix of relationships between them. Softmax(·) represents the normalized exponential operation performed along the allowed node dimension, and d represents the feature dimension. The set of allowed off-diagonal node pairs Only four candidate conversion paths—electricity-cooling, electricity-heating, gas-heating, and gas-electric—were selected based on the actual equipment connections within the park; physical connection mask. diagonal elements and In the algorithm, elements of nodes that are actually connected are set to 0, and all other off-diagonal elements are set to negative infinity; √d represents the attention scaling factor; hydrogen energy and energy storage states are not included. or ; This indicates that the cross-energy correlation characteristics still correspond to the four types of terminal loads along the node axis; The relationship bias matrix formed by historical correlation, equipment operating status, hydrogen energy status, or energy storage status can be represented as:

[0123] ,

[0124] in, This indicates the correlation between load node i and load node j during the historical calibration period; This indicates the start / stop status, load rate, or availability status of the energy conversion equipment connecting the two nodes during time period t; These represent the hydrogen energy state and the energy storage state, respectively. This represents a learnable relational bias mapping.

[0125] The physical connection mask mentioned above Originating from the park's energy system topology, equipment ledger, and input / output port mappings of energy conversion equipment, this constitutes a physical prior hard constraint limiting the permitted connection range; relational bias. Derived from historical correlation and current operating status, it belongs to a learnable dynamic bias that adjusts the strength of the relationship within the allowed connection range. Cannot make the The blocked nodes reconnect.

[0126] The cross-branch adaptive residual fusion unit calculates the weights of the three branches based on time-domain features, frequency-domain features, and cross-energy correlation features on a unified 4-load node axis. The weighted features are then fused and mapped, added to the input embedded feature residuals, and finally normalized to obtain the fused prediction features. This can be represented as:

[0127] ,

[0128] ,

[0129] in, , and These represent the adaptive weights in the time domain, frequency domain, and cross-energy correlation branches, respectively. This indicates the feature pooling operation. and Indicates the parameters for weight calculation. Indicates the fusion mapping parameters, This indicates that layer normalization operations, including mean centering and variance scaling, are performed on each sample feature. These represent time-domain characteristics, frequency-domain characteristics, and cross-energy correlation characteristics, respectively. Indicates fusion prediction features; This is used to normalize the weights of the three branches and make their sum equal to 1.

[0130] The ordered quantile output layer first generates a median prediction result, and then generates low and high quantile prediction results through non-negative offset, which can be expressed as:

[0131] ,

[0132] ,

[0133] ,

[0134] ,

[0135] ,

[0136] ,

[0137] ,

[0138] ,

[0139] in, and These represent the lower quantile level and the upper quantile level, respectively. It is fixed at 0.5 and represents the median level; and These represent the median mapping weight matrix and the median mapping bias vector, respectively. and These represent the lower quantile offset mapping weight matrix and the higher quantile offset mapping weight matrix, respectively. and These represent the low quantile offset mapping bias vector and the high quantile offset mapping bias vector, respectively. and These represent the lower quantile offset and the higher quantile offset, respectively. Softplus(·) is used to ensure that the offsets are non-negative. This represents the multi-energy load forecast range. The above parameterization ensures that the low, medium, and high quantile forecast results are ordered and avoids quantile overlap. ŷ represents the corresponding quantile forecast result.

[0140] During training, the model of this invention divides historical samples into training, validation, and test sets in chronological order, and uses a sliding sample window to construct the input and prediction target. The training set is used to learn embedding parameters, convolution parameters, frequency band parameters, relational attention parameters, fusion weights, and quantile output parameters. The validation set is used to select the model structure and hyperparameters, and the test set is used to evaluate prediction accuracy, interval coverage, and interval width.

[0141] The low, middle, and high quantiles are all used in the training process, and the average quantile loss satisfies the following:

[0142] ,

[0143] ,

[0144] ,

[0145] in, This represents the average quantile loss. Let Q represent the number of training sample windows, M represent the number of load types, H represent the prediction time domain length, and Q = { , , } represents the set of low, middle, and high quantile levels, and |Q| represents the number of elements in set Q. This represents the loss function for the check corresponding to the quantile level q. This represents the prediction error of load type m in the nth sample window in the hth time period. and Let n, m, h, and q represent the actual load value and the quantile forecast value, respectively. In the formula, n, m, h, and q represent the sample window index, load type index, forecast step size index, and quantile level, respectively; max(·) means taking the larger value among the independent variables.

[0146] like Figure 5 As shown, the inputs to the carbon trend calculation module include the terminal multi-energy load quantile prediction results, park boundary definition, external energy supply carbon emission factors, equipment conversion efficiency, baseline operating scheme, and hydrogen energy and energy storage status and energy changes. The carbon trend calculation module takes... The corresponding median forecast results are used as the terminal demand for electricity, cooling, heating, and direct gas.

[0147] The carbon trend calculation module determines the baseline usage of grid-purchased electricity, externally purchased gas, and externally purchased hydrogen based on the balance of electricity, cooling, heating, gas, and hydrogen energy. Specifically, it sets... , , and The median forecasts for terminal electrical load, terminal cooling load, terminal heating load, and direct gas load are represented respectively; all gas quantities and hydrogen energy are converted to energy equivalent rates consistent with power.

[0148] In the benchmark operation plan Under these conditions, the power balance is satisfied:

[0149]

[0150] in, Indicates the power purchased by the power grid. This indicates the local photovoltaic power output. This indicates the gas input equivalent rate of the gas-fired combined heat and power (CHP) equipment. This indicates its power generation efficiency. and These represent the energy storage discharge power and the charging power, respectively. , and These represent the input power of the chiller, heat pump, and electrolytic cell, respectively. Terminal electrical load. This excludes the input from the aforementioned conversion equipment and the energy storage charging power. In the formula, t represents the time period index in the unified time grid, and the superscript 0 indicates a variable under the baseline operating scheme.

[0151] The cooling balance, heating balance, and gas balance respectively satisfy the following:

[0152] ,

[0153] ,

[0154] ,

[0155] in, and These represent the coefficients of performance for the chiller and heat pump, respectively. and These represent the cold storage release rate and the cold charge rate, respectively. Indicates the efficiency of a gas-fired boiler. Indicates the input equivalent rate of the gas-fired boiler. This indicates the heating efficiency of the gas-fired combined heat and power (CHP) equipment. and These represent the heat release rate and the heat charge rate, respectively. Indicates the equivalent rate of purchased natural gas. Direct gas load. This excludes the gas input from gas-fired boilers and gas-fired combined heat and power (CHP) equipment. In the formula, , and These represent the median forecasts for terminal cooling load, terminal heating load, and direct gas load, respectively.

[0156] Hydrogen energy balance is satisfied:

[0157] ,

[0158] in, Indicates the purchased hydrogen equivalent rate. This indicates the efficiency of hydrogen production through electrolysis. This indicates the input power of the electrolyzer under the baseline operating scheme. This indicates the hydrogen energy utilization rate determined by the current operating plan. and These represent the hydrogen release rate and hydrogen replenishment rate, respectively. If the system does not include cold storage, heat storage, gas-fired combined heat and power, or purchased hydrogen, the corresponding variables are set to zero; other energy conversion equipment is added to the corresponding balance equations according to its actual input-output relationship.

[0159] The timing constraints of the stored charge's electrical state satisfy:

[0160] ,

[0161] in, Indicates the electrical state of stored charge. and These represent charging efficiency and discharging efficiency, respectively. The value represents the rated energy storage capacity, and Δt represents the time length between adjacent unified time points; the state changes of cold storage, heat storage, and hydrogen storage are established using similar equations based on the corresponding charge / discharge energy, efficiency, and capacity. Indicates the reference charging power. This indicates the reference discharge power.

[0162] The baseline consumption of electricity, natural gas, and hydrogen purchased from the grid, which cross the park boundary, is directly obtained from the above-mentioned energy balance. For each type of external energy supply that crosses the park boundary, the baseline consumption is multiplied by the corresponding carbon emission factor to obtain the carbon emission contribution of the energy supply. These contributions are then aggregated by time period to form a comprehensive carbon emission trend.

[0163] ,

[0164] ,

[0165] ,

[0166] ,

[0167] in, , and These represent the carbon emission factors for electricity purchased from the grid, natural gas purchased from outside sources, and hydrogen purchased from outside sources, respectively. , and These represent the carbon emission contributions of three types of external energy sources. This represents the overall carbon emission trend value.

[0168] Cold, heat, energy storage, and self-produced hydrogen only participate in the above balance and are not listed as separate carbon emission items alongside grid-purchased electricity, purchased natural gas, or purchased hydrogen; local photovoltaic power is only used as an internal supply item within the park to reduce grid-purchased electricity in the power balance and is not listed as an independent carbon emission item. Where... , and These represent the power purchased from the grid, the equivalent rate of purchased natural gas, and the equivalent rate of purchased hydrogen across the park boundary under the baseline operating scheme, respectively.

[0169] The inputs and outputs of chillers, heat pumps, boilers, gas-fired combined heat and power equipment, and electrolytic cells are respectively entered into the corresponding balance equations, and the energy storage charging and discharging are entered into the supply term, demand term, and state transition equations; the positive and negative directions of each balance variable and the conversion efficiency adopt the unified conventions disclosed in the invention.

[0170] Terminal cooling load, terminal heating load, energy storage, and self-produced hydrogen are not considered additional carbon emission items; local photovoltaic power only reduces the amount of electricity purchased from the grid through power balancing and is not listed as an independent carbon emission item. Therefore, even if the electricity used for chillers or the charging of energy storage is already included in the purchase of electricity from the grid, it is only counted once in the purchase of electricity from the grid.

[0171] like Figure 6 As shown, the low-carbon recommendation module uses a normalization function with fixed upper and lower bounds and out-of-bounds truncation rules to calculate the recommendation urgency and prediction reliability. The urgency threshold is determined based on a preset high-urgency trigger ratio, and the reliability threshold is determined based on the allowed low-reliability ratio, and is fixed within the current application cycle. The target interval coverage is only used for calibration of the low and high quantile prediction intervals and is not used to determine the reliability threshold. Upper-level low-carbon economy optimization is only initiated when the recommendation urgency and prediction reliability reach their respective thresholds.

[0172] Specifically, the normalization function is used to account for load, carbon trend, and interval uncertainty, and its expression is:

[0173] ,

[0174] in, This represents the truncation normalization function based on the calibration dataset D; the calibration dataset D is an independent calibration set other than the training data, or a dedicated calibration subset divided from the validation set, and is not mixed with the test set; and These represent the 5th and 95th percentiles of the calibration dataset, respectively, or the lower and upper bounds predetermined by the engineering allowable range, which is determined by equipment nameplate limits, energy contract boundaries, operating procedures, or operator-preset allowable intervals; when Below or higher The time is truncated to 0 or 1 respectively. This indicates a positive number that prevents the denominator from being zero. and The upper and lower bounds and the threshold are determined during model deployment or periodic calibration; the current application cycle is the running cycle from the completion of one model deployment or threshold calibration to the next model redeployment or threshold recalibration, and the upper and lower bounds and the threshold are fixed within the current application cycle. LayerNorm feature normalization within the neural network does not use this formula.

[0175] The formulas for calculating the urgency of the recommendations and the reliability of the predictions are as follows:

[0176] ,

[0177] ,

[0178] ,

[0179] ,

[0180] ,

[0181] ,

[0182] in, This represents the forecast uncertainty resulting from the weighting of the relative widths of the load forecast intervals for each terminal. , and These represent the low, medium, and high quantile prediction results for load type m, respectively. This indicates non-negative load weights whose sum is 1; Indicates the predicted load level. and These represent the normalized carbon trend and the normalized uncertainty, respectively. Indicate the urgency of the suggestion. Indicates the confidence level of the prediction. and The weights are non-negative and their sum is 1.

[0183] The urgency threshold Take the first R in the calibration dataset quantiles Credibility threshold Take the first G from the calibration dataset quantiles ,in This indicates the preset high urgency trigger ratio. This indicates the proportion of low confidence levels that are allowed to be filtered. The aforementioned quantile settings ensure that optimization is triggered only when the urgency is in the higher quantile range of the calibration set and the confidence level does not fall into the low confidence filtering range. This controls both the optimization trigger frequency and the false triggering caused by low confidence predictions, i.e., the urgency threshold. and credibility threshold Fixed within the current application lifecycle, only when and The upper-level low-carbon economic optimization scheduling is triggered at certain times, which can be specifically expressed as:

[0184] ,

[0185] ,

[0186] .

[0187] The upper-level low-carbon economy optimization uses the transferable load adjustment, energy storage charging and discharging power, output of chillers, heat pumps, boilers, electrolytic cells and gas-fired combined heat and power equipment, and the start-up and shutdown status of flexible equipment as decision variables. , denoted as:

[0188] ,

[0189] in, This represents the transferable load adjustment amount for load type m. and These represent the energy storage charging power and discharging power, respectively. This indicates the output or input power of energy conversion device i. Indicates the device's start / stop status. This indicates the optimized time domain length.

[0190] The flexible equipment is equipment whose power, start / stop status, or operating period is adjustable and still meets the energy supply service constraints after adjustment, including chillers, heat pumps, and electrolytic cells.

[0191] Dimensionless operating cost relative to the baseline operating scheme and carbon emissions from external energy supply The optimization goal is to reduce energy consumption in a coordinated manner, while satisfying five categories of energy balance, equipment output and ramp-up, energy storage status and transferable load conservation, and allowable time constraints.

[0192] Specifically, the dimensionless operating cost and carbon emissions from external energy supply It can be represented as:

[0193]

[0194] ,

[0195] in, , and These represent the prices of electricity purchased from the grid, natural gas purchased from outside the grid, and hydrogen purchased from outside the grid, respectively, for time period t. This represents the unit output operating and maintenance cost of device i. This represents the cost of starting device i per operation. This represents a device startup event variable. , and This represents the external energy consumption rate under optimized decision-making; , and The meaning is the same as that for carbon trend calculation. Each summation is performed within the corresponding optimization time domain or device set.

[0196] Based on the baseline operating plan The operating costs and absolute carbon emissions under these conditions are compared with Using the larger of the two values ​​as the normalized denominator, construct a collaborative optimization objective:

[0197] ,

[0198] ,

[0199] ,

[0200] in, and The weights are non-negative objective weights and their sum is 1. Ω represents the optimal scheduling result, and Ω represents the feasible region consisting of the energy balance and the following equipment operation constraints. This baseline normalization does not require the normalization result to be between 0 and 1; after optimization, the external energy consumption is recalculated based on u, and the above baseline carbon trend calculation is not performed by using the optimized energy supply path. represents a positive number to prevent the normalization denominator from being zero; argmin(·) represents the decision scheme that minimizes the weighted objective within the parentheses within the feasible region Ω.

[0201] During the optimization process, the above five types of energy balance equations remain valid, and the baseline variables with superscript 0 are replaced with optimization variables; the constraints on equipment output, ramp-up, and start-up events are satisfied:

[0202]

[0203]

[0204]

[0205] in, and Let represent the minimum and maximum output parameters of device i, respectively. and Let these represent the unit time limits for decreasing and increasing output power of device i, respectively. Indicates the binary running state. This represents the binary startup event variable.

[0206] Energy storage state meets:

[0207]

[0208]

[0209]

[0210]

[0211] in, and These represent the lower and upper limits of the energy storage state of charge, respectively. and These represent the maximum charging power and the maximum discharging power, respectively; simultaneous charging and discharging within the same time period are avoided through mutually exclusive state variables for charging and discharging or equivalent mutually exclusive constraints. The subscripts in the formula correspond to: t represents the optimization time period index, es represents energy storage, ch and dis represent charging and discharging, respectively, and min and max represent the lower and upper limits of the allowable limit, respectively.

[0212] Transferable loads satisfy conservation, expressed as:

[0213]

[0214]

[0215] in, and Let H and t represent the minimum and maximum allowable transfer adjustment amounts for load type m during time period t, respectively. Both are zero during periods when transfer is not permitted. The adjustment conservation formula ensures that only the energy consumption period changes within the prediction time domain without altering the total energy of this type of transferable load. In the formula, H represents the optimal time domain length, and the summation expression represents the conservation of transferable load adjustment energy within the prediction time domain.

[0216] The external energy consumption is calculated according to the baseline variables and the optimized variables before and after optimization, respectively, without cross-referencing. The solver can use linear programming, mixed integer linear programming, quadratic programming, or other optimization algorithms suitable for the field model.

[0217] The optimization result analysis rules include: pairing negative and positive periods of transferable load adjustment to form load transfer suggestions; generating energy storage charging and discharging suggestions based on energy storage charging and discharging power; mapping the output of energy supply paths or energy conversion equipment to energy conversion suggestions; mapping changes in the start-up and shutdown of flexible equipment to equipment operation adjustment suggestions; and constructing a set of control suggestions from the above suggestions. Carbon signal push suggestions are generated based on high-carbon periods, low-carbon periods, and the urgency of suggestions in the overall carbon emission trend. These suggestions are not directly analyzed from the optimal scheduling variables.

[0218] For each regulation suggestion in the regulation suggestion set, a single-suggestion counterfactual recalculation is performed. The marginal contribution is determined by the difference in carbon emissions and costs between the counterfactual scheduling result without that suggestion and the optimal scheduling result. The construction process and definition of the marginal contribution are as follows:

[0219] Optimal scheduling result The counterfactual feasible region after canceling suggestion j and the counterfactual scheduling result respectively satisfy:

[0220]

[0221]

[0222]

[0223] in, Denotes the feasible region of the original optimization problem. This represents the weighted optimization objective based on the aforementioned operating costs and carbon emissions. This represents the set of decision variables that form the j-th regulatory recommendation. Indicates the baseline operating plan, This represents the counterfactual feasible region formed after fixing the variables in the set of decision variables to baseline values, and argmin(·) represents the decision scheme that minimizes the objective function within the corresponding feasible region. Therefore, and The difference in carbon emissions and costs between the recommendations can represent the marginal change caused by canceling recommendation j, and serve as the basis for scoring subsequent recommendations.

[0224] The aforementioned marginal contribution is defined as:

[0225]

[0226]

[0227] The marginal contributions mentioned above are only used for suggestion ranking and are not required to be equal to the total optimization benefit when the sum of the marginal contributions of each suggestion is equal to the total optimization benefit.

[0228] The initial score for regulatory recommendations considers expected carbon emission reductions, time-sensitive urgency, feasibility of implementation, and positive cost increments. The initial score for carbon signal push recommendations only considers time-sensitive urgency and information push feasibility.

[0229] The original scores for the regulation recommendations and carbon signal push recommendations are as follows:

[0230]

[0231]

[0232]

[0233]

[0234]

[0235]

[0236] in, This represents the set of suggestions for load transfer, energy storage charging and discharging, energy conversion, and equipment operation adjustments resulting from changes in optimal scheduling variables. This represents a set of carbon signal recommendations based on comprehensive carbon emission trends. and These represent the original scores for regulatory recommendations and signal recommendations, respectively; for , This represents the difference between carbon emissions without the recommendation and the optimal carbon emissions. This represents the difference between the cost of the optimal outcome and the cost without the suggestion; a positive value indicates that the suggestion increases the cost. No direct carbon emission reduction or cost increment items are set for changes in scheduling variables; For each scoring indicator, truncation normalization is performed using the fixed upper and lower bounds of the corresponding indicator in the calibration dataset D; and These represent the control suggestion score sample set and the signal suggestion score sample set obtained by running the suggestion generation process on the calibration dataset D, respectively. and By using the fixed upper and lower bounds of the two types of rating sample sets mentioned above, the corresponding original ratings are truncated and normalized to 0 to 1, resulting in recommendation priorities that can be compared on the same scale. ; Indicates the suggested implementation period. The value ranges from 0 to 1, indicating the feasibility of execution based on the device's availability, control permissions, permitted time periods, or the availability of information push channels. The control permissions include automatic control, manual confirmation, or prompt-only permissions configured by the external scheduling control system for the device. to as well as , Non-negative weights; all recommendations are to follow Sort uniformly from highest to lowest. This only indicates the urgency of the time period. This only indicates the priority of specific recommendations; the two should not be used interchangeably. This indicates the weight of the k-th scoring indicator in the regulatory recommendations, where k ranges from 1 to 4; , These represent the weights of time-based urgency and information dissemination feasibility in carbon signal push recommendations, respectively. The weights are preset by the operator based on the operational strategy of prioritizing emission reduction, cost, or execution, or by performing historical replay on the candidate weight combinations on the calibration dataset D to select the weight combination that provides the best comprehensive evaluation of the carbon emission reduction contribution, execution feasibility, and cost increment of the preset number of recommendations before sorting. Once the weights are determined, they remain fixed within the current application cycle.

[0237] The output of the low-carbon recommendation module includes at least the recommendation priority, recommendation type, recommendation target, recommendation execution period, adjustment direction, and recommendation content, and is sent to the external scheduling control system for execution. The external scheduling control system serves only as an execution framework, and this invention does not limit its specific control protocol, actuator type, or underlying control algorithm.

[0238] In another embodiment, the uniform time granularity can be five minutes, ten minutes, thirty minutes or one hour, and the prediction time domain can be set according to the device response time and business needs; the number of frequency bands can be more than three; the carbon trend can adopt a single central trend, or it can be extended to a carbon emission range based on low and high quantile results, as specified in the specification.

[0239] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-energy load forecasting and low-carbon optimization control, characterized in that, Includes the following steps: S1 acquires terminal multi-energy load, meteorological and time information, equipment operating status and time-varying data, as well as park boundary and carbon accounting parameters; S2 performs anomaly identification, missing data filling, timestamp correction, and unified time grid alignment on the data obtained from S1 to obtain a unified load sequence sample composed of terminal electricity, cooling, heating loads, and direct gas loads, exogenous features, and external energy supply carbon emission factors and time-of-use energy prices aligned by time period; S3, perform multi-frequency decomposition on the unified load sequence sample to obtain four types of components: trend component, periodic component, pulse component and high-frequency fluctuation component. S4. Input the four types of components, exogenous features, time location codes and load type codes into the multi-energy load interval prediction model. After parallel extraction of embedding splicing, time domain dependent features, adaptive frequency domain features and multi-energy relationship features and adaptive residual fusion, the median prediction sequence and multi-energy load prediction interval are obtained. S5. Using the median prediction sequence as the terminal demand benchmark, combined with the park boundary and carbon accounting parameters, and according to the balance of electricity, cooling, heating, gas and hydrogen energy, the external energy supply benchmark consumption that crosses the park boundary is determined and multiplied by the corresponding carbon emission factor to obtain the carbon emission contribution. Then, the carbon emission trend is summarized by time period. S6. Based on the multi-energy load forecast range and the comprehensive carbon emission trend, generate low-carbon recommendations and send them to the external dispatch control system for execution.

2. The multi-energy load forecasting and low-carbon optimization control method as described in claim 1, characterized in that, In S2, anomaly identification is: identifying data that exceeds limits, has long-term constant values, has repeated timestamps, and has anomalous mutations, and marking them as invalid observations; Missing data filling involves filling in the original missing data and data gaps formed by invalid observations. Short-term gaps are filled using linear interpolation, while long-term gaps are filled using historical data from similar periods of the same energy type. Timestamp correction involves adjusting the timestamp based on the sampling period, reporting delay, and device clock offset. Unified time grid alignment is as follows: a unified time grid is constructed with a preset time granularity, and the corrected data is aggregated, interpolated, or resampled to the unified time grid to obtain a unified load sequence sample composed of terminal electrical load, terminal cooling load, terminal heating load, and direct gas load.

3. The multi-energy load forecasting and low-carbon optimization control method as described in claim 1, characterized in that, Specifically, S3 includes: independently performing trend extraction, period identification and reconstruction, pulse identification, and high-frequency margin separation on the terminal electrical load, terminal cooling load, terminal heating load, and direct gas load sequences; The trend extraction method is as follows: trend components are extracted by averaging through a sliding window. The periodic identification and reconstruction are as follows: performing fast Fourier transform, power spectral density calculation, Top-K positive and negative frequency conjugate peak pair screening, and fast Fourier inverse transform on the residual components after removing the trend components to reconstruct the periodic components; The pulse identification and high-frequency residual separation are performed as follows: robust threshold identification based on the median absolute deviation is performed on the remaining residual after removing the periodic component, the pulse component is extracted and removed, and continuous high-frequency fluctuation components are obtained.

4. The multi-energy load forecasting and low-carbon optimization control method as described in claim 1, characterized in that, The S4 multi-energy load interval prediction model includes: an input embedding layer, a time-domain dependent feature extraction unit, an adaptive frequency-domain feature extraction unit, a multi-energy relationship feature extraction unit, a cross-branch adaptive residual fusion unit, and an ordered quantile output layer; The input embedding layer maps the trend component, periodic component, pulse component and high-frequency fluctuation component to a unified feature dimension, and then concatenates the mapped exogenous features with the mapping results of the four types of load components, time position code and load type code along the channel dimension to obtain the embedding features with terminal electrical load, terminal cooling load, terminal heating load and direct gas load as the four load nodes. The embedded features are processed in parallel into three branches: a time-domain dependent feature extraction unit, an adaptive frequency-domain feature extraction unit, and a multi-energy relationship feature extraction unit, which respectively obtain time-domain features, frequency-domain features, and cross-energy correlation features, and all three branches maintain the same 4-load node axis; The cross-branch adaptive residual fusion unit calculates the weights of three branches based on time-domain features, frequency-domain features, and cross-energy correlation features on a unified 4-load node axis. It then performs fusion mapping on the weighted branch features and adds them to the input embedded feature residuals. Finally, it obtains the fused prediction features through normalization. The ordered quantile output layer inputs fused prediction features, first forming a median prediction result, and then forming low quantile and high quantile prediction results through non-negative offset, such that the low quantile prediction result is smaller than the median prediction result and the median prediction result is smaller than the high quantile prediction result.

5. The multi-energy load forecasting and low-carbon optimization control method as described in claim 4, characterized in that, The temporal-dependent feature extraction unit employs dilated causal convolution, which ensures that the output of time period t does not use observations after time period t. The dilation rate increases with the number of layers to expand the temporal receptive field. The convolution output is then nonlinearly activated and added to the embedded feature residual to obtain the temporal features.

6. The multi-energy load forecasting and low-carbon optimization control method as described in claim 4, characterized in that, The adaptive frequency domain feature extraction unit includes a frequency domain main path and a pulse retention bypass; In the frequency domain main path, the embedded features are first transformed to the frequency domain by Fast Fourier Transform (FFT). Then, the low-frequency, mid-frequency, and high-frequency bands are divided by learnable frequency band boundaries or learnable masks, and trend enhancement, period enhancement, and useful high-frequency fluctuation selection are performed respectively. After that, the inverse FFT of each frequency band is performed. In the inverse FFT, conjugate symmetric constraints are applied to the frequency weights so that the low-frequency, mid-frequency, and high-frequency features obtained by the inverse transform are all real-valued time-domain features. The pulse retention bypass directly reads the channel formed by independent mapping of pulse components, and retains abrupt change information through pulse gating to obtain pulse characteristics; The time-domain features of the low-frequency, mid-frequency, and high-frequency bands after inverse transformation are spliced ​​and adaptively fused with the pulse features, and then added to the residual of the original embedded features to obtain the frequency domain features.

7. The multi-energy load forecasting and low-carbon optimization control method as described in claim 4, characterized in that, The multi-energy relationship feature extraction unit constructs a multi-energy relationship graph based on the embedding features of four load nodes corresponding to terminal electrical load, terminal cooling load, terminal heating load, and direct gas load, as well as the physical connection relationships of the equipment. The physical connection mask only allows actual electric-cooling, electric-heating, gas-heating, and gas-electric conversion paths between the four types of load nodes. The diagonal elements of the mask and the allowed node pairs are set to zero, and the remaining off-diagonal elements are set to negative infinity. Between the allowed node pairs, a relationship bias is generated based on historical correlation, the start / stop status of energy conversion equipment, load rate or availability status, and hydrogen energy and energy storage status. The embedding features of the nodes are mapped to queries, keys, and values. The relationship weights between nodes are calculated, and the physical connection mask and relationship bias are superimposed before normalization to obtain the final relationship weight matrix. Then, graph attention aggregation is performed to output the cross-energy association features corresponding to the four types of loads.

8. The multi-energy load forecasting and low-carbon optimization control method as described in claim 1, characterized in that, Specifically, S5 includes: taking the median forecast results of terminal electrical load, terminal cooling load, terminal heating load and direct gas load as the terminal demand benchmark; Five types of energy balance are established under the baseline operation plan: power balance, cooling balance, heating balance, gas balance and hydrogen energy balance; the baseline operation plan is the currently confirmed operation plan. When there is no currently confirmed operation plan, it is composed of preset energy supply paths and corresponding equipment output, energy storage status and external energy consumption. In the aforementioned power balance, grid-purchased electricity, local photovoltaic power, gas-fired combined heat and power generation, and energy storage and discharge are considered as supply items, while terminal electrical load, chillers, heat pumps, electrolyzers, and energy storage and charging are considered as demand items. In the aforementioned cooling balance, the cooling supply from the chiller and the cooling release from the cold storage meet the terminal cooling load and the cooling charge from the cold storage. In the heat balance, gas boilers, heat pumps, gas combined heat and power, and thermal storage heat release meet the terminal heat load and thermal storage heat charging. In the aforementioned gas balance, the purchased gas meets the gas input requirements for direct gas load, gas boilers, and gas-fired combined heat and power. In the aforementioned hydrogen energy balance, purchased hydrogen, electrolytic hydrogen production, and hydrogen storage and release meet the hydrogen energy usage and hydrogen storage and charging requirements. The baseline consumption of electricity purchased from the grid, gas purchased from outside, and hydrogen purchased from outside the park is directly obtained from the above five categories of energy balance. For each type of external energy supply that crosses the park boundary, the baseline consumption is multiplied by the corresponding carbon emission factor to obtain the carbon emission contribution of the energy supply, and the results are summarized by time period to form a comprehensive carbon emission trend.

9. The multi-energy load forecasting and low-carbon optimization control method as described in claim 8, characterized in that, The low-carbon recommendation generation in S6 specifically includes: Based on a calibration dataset, the suggestion urgency and prediction confidence are obtained after normalization with fixed upper and lower bounds and out-of-bounds truncation. The urgency threshold is determined according to the preset high urgency trigger ratio, and the confidence threshold is determined according to the allowed low confidence ratio, and is fixed within the current application cycle. Upper-level low-carbon economy optimization is triggered only when the urgency of the recommendation reaches the urgency threshold and the confidence of the prediction reaches the confidence threshold. The upper-level low-carbon economic optimization takes the transferable load adjustment, energy storage charging and discharging power, energy conversion equipment output and flexible equipment start-up and shutdown status as decision variables, and takes the dimensionless operating cost relative to the benchmark operation plan and the coordinated reduction of carbon emissions from the external energy supply side as optimization objectives, while satisfying five categories of energy balance, equipment output and ramp-up, energy storage status, transferable load conservation and allowable time period constraints. The optimization results are analyzed and combined with the comprehensive carbon trend to generate a set of regulation suggestions and carbon signal push suggestions. The set of regulation suggestions includes load transfer suggestions, energy storage charging and discharging suggestions, energy conversion suggestions, and equipment operation adjustment suggestions. The carbon signal push suggestions are generated based on the high carbon period, low carbon period, and urgency of the suggestions in the comprehensive carbon emission trend.

10. The multi-energy load forecasting and low-carbon optimization control method as described in claim 9, characterized in that, For each regulation suggestion in the set of regulation suggestions, a single suggestion counterfactual recalculation is performed, and the marginal contribution is determined by the difference in carbon emissions and cost between the counterfactual scheduling result without the suggestion and the optimal scheduling result; The original scores for the control recommendations are calculated based on the expected carbon emission reduction, time urgency, implementation feasibility, and positive cost increment. The original scores for the carbon signal push recommendations are calculated based on the time urgency and information push feasibility. The two types of original scores are truncated and normalized to a uniform scale and then arranged in descending order to obtain the priority of the recommendations.

Images