A cold chain refrigeration equipment energy consumption intelligent scheduling and cost control system

Abstract

The application is particularly a cold chain refrigeration equipment energy consumption intelligent scheduling and cost control system, relating to the technical field of cold chain logistics energy management, comprising: a probability prediction module, used for respectively predicting the refrigeration load demand and the electricity price on multiple time scales based on the unified data space-time tensor, outputting the load prediction interval and the electricity price prediction interval of each period, and calculating the confidence score of each prediction result. In the application, the cold storage energy storage device is taken as a flexible load buffer element into the collaborative scheduling, the cold storage energy storage device is instructed to charge cold in the valley electricity period, and the stored cold is preferentially released to meet the cold chain load demand in the peak electricity period, so that the refrigeration electricity load originally concentrated in the peak electricity period is transferred to the valley electricity period for execution, and the total electricity cost is reduced by using the peak-valley electricity price difference.

Description

Technical Field

[0001] This invention relates to the field of energy management technology in cold chain logistics, and in particular to an intelligent scheduling and cost control system for energy consumption of cold chain refrigeration equipment. Background Technology

[0002] In cold chain logistics operations, the electricity consumption of refrigeration systems accounts for a significant proportion of total operating costs. Existing cold chain refrigeration systems typically operate at constant power or use temperature threshold-based on-off control. Refrigeration units respond to cold chain loads on demand, without incorporating cold storage energy as a flexible load buffer in time-series scheduling. Under this operating mode, refrigeration electricity load is concentrated during peak hours, making it impossible to utilize the peak-valley price difference of time-of-use electricity to shift the load to off-peak hours, resulting in persistently high electricity costs.

[0003] Furthermore, traditional load forecasting methods often employ single-point forecasting, outputting only the predicted load value without quantifying the uncertainty of the forecast. Scheduling decisions lack input for assessing the risk of forecast errors, which can easily lead to scheduling strategy failure or cold storage temperature exceeding limits when forecast deviations are large. Cold chain refrigeration equipment typically operates only as a passive electrical load, without coordinated scheduling with distributed energy sources such as photovoltaic power generation and energy storage batteries, nor participating in demand response interactions on the grid side. Therefore, there is room for improvement in energy utilization efficiency and economic benefits. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent scheduling and cost control system for energy consumption of cold chain refrigeration equipment in order to solve the above-mentioned problems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A smart energy consumption scheduling and cost control system for cold chain refrigeration equipment includes: The data acquisition and fusion module is used to collect operating data of refrigeration units and cold storage energy storage devices in the cold chain refrigeration system, as well as time-of-use electricity price data. After time alignment and standardization processing, a unified data spatiotemporal tensor is generated. The probability prediction module is used to perform probability predictions on cooling load demand and electricity price at multiple time scales based on the unified data spatiotemporal tensor, output the load prediction interval and electricity price prediction interval for each time period, and calculate the confidence score of each prediction result. The collaborative scheduling optimization module is used to take the load prediction interval, the electricity price prediction interval, the confidence score and the current cold storage capacity of the cold storage energy storage device as inputs, with the minimization of total electricity cost as the primary optimization objective, and to perform multi-objective optimization under the hard constraint that the cold storage temperature does not exceed the safety threshold. It outputs the operating frequency of the refrigeration unit and the charging power or releasing power of the cold storage energy storage device in each time period, thus forming a collaborative scheduling strategy. The integrated scheduling execution module is used to dynamically switch between multiple energy supply modes based on the aforementioned collaborative scheduling strategy, combined with photovoltaic power generation and grid demand response instructions, output integrated scheduling instructions, and issue them for execution.

[0006] Preferably, the step of collecting operating data and time-of-use electricity price data of refrigeration units and cold storage devices in the cold chain refrigeration system, and generating a unified data spatiotemporal tensor after time alignment and standardization processing, further includes: On the edge computing node, the raw sensor data is subjected to moving average filtering noise reduction processing, and the length of the filtering window is configured according to the noise characteristics of each sensor signal. The mean, variance and rate of change in each sampling window are extracted from the raw time series data as feature vectors, and the feature vectors are used to replace the raw high-frequency data and uploaded to the cloud computing platform. The time alignment process employs a linear interpolation algorithm to uniformly interpolate sensor data with different sampling frequencies to the same time resolution. Z-score standardization is performed on various types of sensor data to eliminate the dimensional differences between different physical quantities, and then the data is organized into the unified data spatiotemporal tensor according to the time dimension and the spatial dimension of each sensor node.

[0007] Preferably, the probabilistic prediction of cooling load demand at multiple time scales includes: Cold chain storage data, cargo entry and exit plan data, historical outdoor temperature and humidity data, and weather forecast information are input into a large time series model. The large time series model performs feature encoding and sequence inference on the input multivariate time series data, and outputs the cold load probability distribution for each time period at three time granularities: day-ahead scale, intraday scale, and real-time scale. The upper quantile and lower quantile corresponding to the preset confidence level are extracted from the cold load probability distribution for each time period to form the load prediction interval for each time period. The feature vectors of each sensor node in the unified data spatiotemporal tensor are input into a graph neural network. The graph neural network forms a graph structure with each sensor node as a vertex and the physical connection relationship between nodes as edges. The features of adjacent nodes are aggregated through graph convolutional layers to output the feature representations of each node that fuse spatial correlation information. The feature representations of each node are then concatenated with the time series features output by the time series large model and re-input into the prediction head of the time series large model to output the updated probability distribution of cooling load for each time period.

[0008] Preferably, the probabilistic prediction of electricity prices includes: Using historical time-of-use electricity price data and power grid dispatch announcements as input, a gradient boosting regression algorithm is employed to predict electricity prices for each future time period. The inputs to the gradient boosting regression algorithm are historical electricity price sequences, date type identifiers, and key fields from power grid dispatch announcements, and the output is the expected electricity price value for each future time period. Based on the prediction residual distribution of the gradient boosting regression algorithm, the upper and lower bounds of the electricity price prediction for each time period are calculated, forming the electricity price prediction interval. The system analyzes demand response event notifications issued by the grid side, extracts the response time window and response demand power parameters, and calculates the trigger probability of demand response events in future time periods based on the trigger frequency of historical demand response events and the current grid load status.

[0009] Preferably, the confidence score is calculated as follows: Based on the backtesting error statistics of each prediction model on recent historical data, a confidence score is calculated. The confidence score is equal to 1 minus the average absolute value of the relative error between the predicted value and the actual value in each backtesting period within a recent preset time window. Wherein, when the actual value of a certain backtesting period is zero, that period is not included in the error statistics. The confidence score ranges from 0 to 1. The larger the value, the higher the confidence level of the prediction. The confidence score is used as the weight coefficient of each prediction input in the subsequent multi-objective optimization solution.

[0010] Preferably, the multi-objective optimization solution employs a security-constrained deep reinforcement learning framework, which consists of a policy network, a value network, and a security constraint layer, wherein: The input to the strategy network is a state vector consisting of the current cold storage capacity, cold storage temperature, current electricity price, load prediction range, and electricity price prediction range, and the output is an action vector consisting of the operating frequency setpoint of each refrigeration unit and the charging and discharging power setpoint of the cold storage energy storage device. The input to the value network is the same as that of the policy network, and the output is the estimated cumulative reward for performing the corresponding action in the current state. The safety constraint layer performs temperature constraint verification on the action vector output by the policy network. The Lagrange relaxation method is used to embed the cold storage temperature safety constraint into the policy update process. When the predicted value of the cold storage temperature approaches the boundary of the cold storage temperature safety threshold, the penalty coefficient of the constraint term is automatically increased, driving the policy away from the constraint violation area.

[0011] Preferably, the reward function of the safety constraint deep reinforcement learning framework is composed of a weighted sum of the total electricity cost term, the load power variance term, and the compressor start-stop penalty term. The three terms are Z-score standardized before being substituted into the reward function calculation to eliminate the difference in dimensions. Among them, the weight coefficient of the total electricity cost item is positively correlated with the confidence score of the electricity price forecast, and the weight coefficients of the load power variance item and the compressor start-stop penalty item are positively correlated with the confidence score of the load forecast. When the confidence score is high, the corresponding weight coefficient of the forecast input increases, and the collaborative scheduling strategy uses the forecast results to perform peak-valley arbitrage. When the confidence score is low, the corresponding weight coefficient decreases, and the collaborative scheduling strategy prioritizes ensuring the safety constraints of the cold storage temperature.

[0012] Preferably, the collaborative scheduling strategy is decomposed and executed at three levels: day-ahead, intraday, and real-time, wherein: At the day-ahead level, based on the day-ahead load forecast interval and electricity price forecast interval, a macro power allocation scheme for each period of the next 24 hours is generated to determine the target operating status of chiller units and cold storage energy storage devices for each period. At the intraday level, every preset rolling cycle, the macroscopic power allocation scheme is rolled over and corrected based on the intraday scale prediction update data and actual operating deviations. At the real-time level, model predictive control is adopted, taking real-time sensor data and real-time scale prediction data as inputs, and outputting fine-tuning values ​​of control quantities within a short future time window to fine-tune the operating frequency of the chiller unit and the charging and discharging power of the cold storage energy storage device in real time.

[0013] Preferably, the dynamic switching between multiple energy supply modes includes: When the photovoltaic power generation is higher than the preset threshold, the photovoltaic power is allocated to the operation of the chiller unit or the cooling energy storage device; when the photovoltaic power generation is lower than the preset threshold, the power is switched to grid power and the operating mode is determined according to the current electricity price signal; during the peak electricity price period, the energy storage battery releases electricity to the chiller unit and operates synchronously with the cooling energy storage device. When a load reduction command is received from the grid, the stored cold energy in the cold storage device is used to replace the immediate cooling output of the refrigeration unit; when a load increase command is received, the operating power of the refrigeration unit is increased or the cold storage device is started to charge the cold energy. The output of the integrated scheduling instruction also includes: collecting actual execution result data, adding the actual operation data within the most recent preset period to the training dataset, and incrementally training and updating the strategy network in the multi-objective optimization solution so that the collaborative scheduling strategy can adapt to the long-term changing trends of electricity price structure and load pattern.

[0014] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention incorporates a cold storage energy storage device as a flexible load buffer element into the coordinated scheduling. During off-peak hours, the cold storage energy storage device is instructed to charge for cooling, and during peak hours, the stored cold energy is released first to meet the cold chain load demand. This transfers the cooling power load that was originally concentrated during peak hours to off-peak hours, and reduces the total electricity cost by utilizing the peak-valley electricity price difference.

[0015] 2. This invention outputs the load forecast interval by using a probabilistic forecasting method in the load forecasting stage, and transmits the forecast uncertainty to the scheduling optimization stage in the form of confidence score. This enables the collaborative scheduling strategy to dynamically adjust the safety margin according to the forecast confidence level, which improves the robustness of scheduling decisions under forecast error conditions compared with single-point forecasting schemes. Attached Figure Description

[0016] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0017] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0018] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0019] Example 1

[0020] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0021] In this embodiment, it includes: In cold chain logistics scenarios, the electricity consumption of refrigeration systems accounts for a significant proportion of total operating costs. Existing cold chain refrigeration systems typically operate in constant power or temperature threshold-based on-off control modes, failing to fully utilize the peak-valley price difference of time-of-use electricity pricing and the time-series adjustment capabilities of cold storage energy storage systems, resulting in persistently high electricity costs. Furthermore, traditional load forecasting methods are mostly single-point forecasts, making it difficult to quantify forecast uncertainties and leaving scheduling decisions without risk assessment input. In addition, cold chain refrigeration equipment operates only as passive electrical loads, failing to coordinate with distributed energy sources such as photovoltaic power generation and energy storage batteries, and not participating in demand response interactions on the grid side, leaving room for improvement in energy utilization efficiency and economic benefits.

[0022] In this embodiment, the cold chain refrigeration system includes at least a refrigeration unit and a cold storage energy storage device. The refrigeration unit includes a compressor, condenser, evaporator, and expansion valve. The cold storage energy storage device includes a cold storage tank or a phase change cold storage material area. Temperature sensors, pressure sensors, flow sensors, and power sensors are deployed at key nodes of the refrigeration unit and the cold storage energy storage device, respectively. Cold storage capacity sensors and temperature distribution sensors are deployed in the cold storage energy storage device area. The system also connects to a light intensity sensor, a photovoltaic power generation sensor, and grid electricity price signal interfaces and demand response signal interfaces. All sensors and signal interfaces are connected to a cloud computing platform through edge computing nodes.

[0023] A method for coordinated scheduling of cold chain refrigeration and cold energy storage according to an embodiment of the present invention includes the following: Step S100: Collect and fuse multi-source heterogeneous operation data of the cold chain refrigeration system to generate a unified data spatiotemporal tensor.

[0024] Operating parameters such as compressor operating frequency, condensing pressure, evaporating temperature, and instantaneous power consumption are collected in real time through temperature sensors, pressure sensors, flow sensors, and power sensors deployed at key nodes of the refrigeration unit. Simultaneously, current cold storage capacity and internal temperature distribution data of the cold storage energy storage device are collected through cold storage capacity sensors and temperature distribution sensors. Time-of-use electricity price data for the current period is also accessed via a grid electricity price signal interface.

[0025] The collected multi-source heterogeneous data were processed by time alignment, missing value imputation and outlier detection. Z-score standardization was performed on various types of sensor data to eliminate the dimensional differences between different physical quantities and generate a unified data spatiotemporal tensor.

[0026] It should be noted that the time alignment process described above refers to the following: The sampling frequencies of the various sensors are different; the sampling period for temperature sensors is on the order of seconds, while the update period for electricity price signals ranges from minutes to hours. The time alignment process uses a linear interpolation algorithm to uniformly interpolate data from different sampling frequencies to the same time resolution, forming a multidimensional data matrix with a consistent time axis. Missing value imputation uses linear interpolation or forward padding to complete missing data during offline periods. Outlier detection marks and removes data points that exceed the physical range of each parameter. The multidimensional data processed in the above way is organized into a unified spatiotemporal tensor according to the time and spatial dimensions (the location of each sensor node).

[0027] In this embodiment, to reduce the data transmission bandwidth and computational load of the cloud computing platform, real-time filtering, noise reduction, and feature extraction operations are performed on the raw sensor data at the edge computing node, and the processed compressed data is then uploaded to the cloud computing platform. The filtering and noise reduction uses a moving average filter, and the filter window length is configured according to the noise characteristics of each sensor signal. The feature extraction operation extracts the mean, variance, and rate of change within each sampling window from the raw time-series data as feature vectors, which are then used to replace the raw high-frequency data for uploading.

[0028] Step S200: Based on the unified data spatiotemporal tensor, perform probabilistic predictions of cooling load demand and electricity price at multiple time scales, and output the load prediction interval and electricity price prediction interval for each time period.

[0029] Using cold chain storage data, cargo entry and exit plan data, historical outdoor temperature and humidity data, and weather forecast information from the unified data spatiotemporal tensor as inputs, the system performs probabilistic predictions of the cooling load demand of the refrigeration system at three time granularities: day-ahead, intraday, and real-time. The system outputs a load prediction interval consisting of the upper and lower bounds of the load prediction for each time period.

[0030] By taking historical time-of-use electricity price data, grid dispatch announcements, and current electricity price signals as inputs, the electricity price for each future time period is predicted, and the expected electricity price for each time period and the peak-valley time period division results are output to obtain the electricity price prediction range.

[0031] Simultaneously, confidence scores are calculated for both the load forecast interval and the electricity price forecast interval. The confidence scores characterize the reliability of each forecast result and serve as weighted input parameters for subsequent scheduling optimization solutions.

[0032] Step S200 may specifically include: Step S201: Perform multi-timescale probability prediction of cooling load demand and output the load prediction range.

[0033] Cold chain storage data, cargo inbound / outbound plan data, historical outdoor temperature and humidity data, and weather forecast information are input into the Time-LLM time series model. The Time-LLM model performs feature encoding and sequence inference on the input multivariate time series data, and outputs the probability distribution of cooling load for each time period at three time granularities: day-ahead scale (next 24 hours), intraday scale (next 4 hours), and real-time scale (next 15 minutes). The upper and lower quantiles corresponding to the preset confidence level are extracted from the probability distribution of cooling load for each time period to form the load forecast interval for each time period.

[0034] It should be noted that the above probabilistic prediction outputs a load forecast interval rather than a single-point forecast value. This is to quantify the uncertainty of load forecasting. In subsequent scheduling optimization solutions, the width of the load forecast interval directly affects the conservatism of the scheduling strategy: when the load forecast interval is wider, the scheduling strategy tends to retain more cooling capacity as a safety margin.

[0035] In this embodiment, to extract the spatial correlation features between various equipment nodes in the cold chain refrigeration system, the feature vectors of each sensor node in the unified data spatiotemporal tensor are input into a graph neural network. The graph neural network constructs a graph structure with each sensor node as a vertex and the physical connections between nodes as edges. The input feature of each node is the feature vector of the corresponding sensor. The graph neural network aggregates the features of adjacent nodes through graph convolutional layers, outputting feature representations of each node that incorporate spatial correlation information. The node feature representations output by the graph neural network are concatenated with the time-series features output by the Time-LLM large-scale time-series model, and then re-input into the prediction header of the Time-LLM large-scale time-series model to output the updated probability distribution of cooling load for each time period. The graph neural network adopts a supervised training mode, using the node-level load prediction error as the loss function, and uses the Adam optimization algorithm for parameter updates.

[0036] Step S202: Predict electricity prices and analyze demand response signals, outputting the electricity price prediction range and the probability of demand response triggering.

[0037] Using historical time-of-use (TOU) electricity price data and grid dispatch announcements as input, a gradient boosting regression algorithm is employed to predict electricity prices for each future time period. The input to the gradient boosting regression algorithm consists of historical electricity price sequences, date type identifiers, and key fields from grid dispatch announcements; the output is the expected electricity price value for each future time period. Based on the prediction residual distribution of the gradient boosting regression algorithm, the upper and lower bounds of the predicted electricity prices for each time period are calculated, forming the electricity price prediction interval.

[0038] Simultaneously, the demand response event notifications issued by the grid side are analyzed to extract the response time window and response demand power parameters. Based on the triggering frequency of historical demand response events and the current grid load status, the triggering probability of demand response events in future time periods is calculated.

[0039] Step S203: Calculate the confidence score for the load forecast results and the electricity price forecast results.

[0040] Confidence scores were calculated for both the load forecast range and the electricity price forecast range. The confidence scores were calculated based on the backtesting error statistics of each forecasting model on recent historical data. The calculation formula is: ; in, This represents the total number of backtesting periods within the recent preset time window. The backtesting period number is used for the backtesting period. For the first The actual value for each backtesting period. For the first The predicted values ​​for each backtesting period. The value ranges from 0 to 1, with a larger value indicating a higher level of prediction reliability. When the actual value for a certain period... During this period, errors are not included in the error statistics to avoid division-to-zero anomalies. The confidence score is then used as a weighting coefficient for each predicted input in the scheduling optimization solution.

[0041] Step S300: Based on the load forecast interval, electricity price forecast interval, and current status data of the cold storage energy storage device, a collaborative scheduling strategy for the chiller unit and the cold storage energy storage device is generated through multi-objective optimization.

[0042] Using load forecast interval, electricity price forecast interval, confidence score, current cold storage capacity of cold storage device and current temperature of cold storage as input, the primary optimization objective is to minimize total electricity cost, while the secondary optimization objectives are to minimize system load power variance and compressor start-stop frequency. Under the hard constraint that the cold storage temperature does not exceed the safety threshold, multi-objective optimization is performed to solve the problem. The output is the start-stop status and operating frequency of refrigeration units in each time period, and the charging or releasing power of cold storage device, which constitutes a collaborative scheduling strategy.

[0043] It should be noted that the aforementioned coordinated dispatch strategy treats cold storage energy storage devices as flexible load buffers. During off-peak electricity hours, the strategy instructs the cold storage energy storage devices to charge at full or near full load, storing the cold energy in storage tanks or phase change storage materials. During peak electricity hours, the strategy instructs the cold storage energy storage devices to prioritize the release of stored cold energy to meet cold chain load demands, while simultaneously reducing or even shutting down the operating frequency of compressors in refrigeration units. Through this timing configuration of charging and releasing cold energy, the refrigeration load, originally concentrated during peak hours, is shifted to off-peak hours, utilizing the peak-valley price difference to reduce overall electricity costs.

[0044] It should be noted that the aforementioned cold storage temperature safety threshold is embedded as a hard constraint in the optimization solution process. The cold storage temperature safety threshold is determined by the type of goods stored; for example, the safe temperature range for frozen goods is -22°C to -18°C, and for refrigerated goods it is 0°C to 4°C. In each iteration of the optimization solution, the current candidate scheduling strategy is checked for temperature constraints. If a candidate strategy causes the predicted cold storage temperature to exceed the cold storage temperature safety threshold at any given time, the strategy is deemed infeasible and eliminated.

[0045] Step S300 may specifically include: Step S301: Configure the objective function and constraints for multi-objective optimization.

[0046] The primary optimization objective is the total electricity cost within the scheduling cycle. The calculation formula is as follows: ; in, This represents the total number of time periods within the scheduling cycle. For time period numbering, For the first Electricity price during specific time periods For the first Total system power consumption during the time period The duration of a single time period.

[0047] Secondary optimization objectives include: system load power variance This is used to measure the smoothness of the load curve; the number of compressor start-stop cycles. This is used to measure the losses caused by frequent equipment start-ups and shutdowns. The hard constraint is that the predicted cold storage temperature for each time period must be within the safe temperature range. The decision variables are the operating frequency of the refrigeration unit and the charging or releasing power of the cold storage energy storage device for each time period.

[0048] Step S302: Use a safety-constrained deep reinforcement learning framework to solve multi-objective optimization and output a cooperative scheduling strategy.

[0049] The objective function, constraints, and decision variables configured in step S301 are input into the safety constraint deep reinforcement learning framework for solution. The safety constraint deep reinforcement learning framework consists of three components: a policy network, a value network, and a safety constraint layer. The policy network's input is a state vector composed of the current cold storage capacity, cold storage temperature, current electricity price, load forecast interval, and electricity price forecast interval. Its output is an action vector composed of the operating frequency setpoints of each refrigeration unit and the charging / discharging power setpoints of the cold storage energy storage device. The operating frequency setpoints in the action vectors are directly used as the target frequency commands for the refrigeration unit's frequency converter, and the charging / discharging power setpoints are directly used as the target power commands for the cold storage energy storage device's power controller. The value network's input is the same as the policy network's, and its output is an estimated cumulative reward for performing the corresponding action in the current state. The safety constraint layer performs temperature constraint verification on the action vectors output by the policy network. A Lagrange relaxation method is used to embed the cold storage temperature safety constraint into the policy update process: when the predicted cold storage temperature approaches the cold storage temperature safety threshold boundary, the penalty coefficient of the constraint term automatically increases, driving the policy away from the constraint violation area.

[0050] reward function The weighted sum of the total electricity cost, load power variance, and compressor start-stop penalty is expressed as: ; in, The total electricity cost for the current scheduling cycle. The system load power variance, This refers to the number of times the compressor starts and stops. , , These are the weight coefficients for the corresponding items. This is the reward function value. Because... , and The three items have different dimensions. Before being substituted into the reward function for calculation, the three items are standardized by Z-score to eliminate the difference in dimensions before being weighted and summed.

[0051] It should be noted that the weights of the above three items are dynamically adjusted based on the confidence score. Electricity price forecast confidence score Positive correlation and With load forecast confidence score Positive correlation: When the confidence score is high, the weight coefficient of the corresponding predicted input increases, and the scheduling strategy more actively utilizes the prediction results for peak-valley arbitrage; when the confidence score is low, the corresponding weight coefficient decreases, and the scheduling strategy tends to be conservative, prioritizing the protection of cold storage temperature safety constraints. The safety-constrained deep reinforcement learning framework adopts an online training mode, using the policy gradient method to update the parameters of the policy network and the value network. After training and convergence in the simulation environment, the agent in the safety-constrained deep reinforcement learning framework outputs the optimal decision variable values ​​for each time period for the current state, constituting a cooperative scheduling strategy.

[0052] In this embodiment, to adapt to scheduling needs at different time granularities, the collaborative scheduling strategy is decomposed and executed at three levels: day-ahead, intraday, and real-time. At the day-ahead level, based on the day-ahead load forecast interval and electricity price forecast interval, a macroscopic power allocation scheme for each time period in the next 24 hours is generated, determining the target operating state of the chiller units and cold storage energy storage devices for each time period. At the intraday level, every preset rolling cycle, based on the intraday forecast update data and actual operating deviations, the day-ahead macroscopic power allocation scheme is rolled over and corrected, adjusting the power allocation values ​​for each time period. At the real-time level, a model predictive control method is adopted. The input of the model predictive control method is the real-time sensor data at the current moment and the real-time forecast data, and the output is the fine-tuning value of the control quantity within a short future time window, which is used to fine-tune the operating frequency of the chiller units and the charging and discharging power of the cold storage energy storage devices in real time to compensate for short-term disturbances.

[0053] Step S400: Based on the collaborative scheduling strategy and distributed energy status data, perform multi-energy collaborative scheduling and grid demand response interaction, and output comprehensive scheduling instructions.

[0054] The system acquires real-time power generation data from the photovoltaic (PV) power generation system and state-of-charge (SOC) data from the energy storage batteries. Based on real-time sunlight conditions and PV power generation, it dynamically adjusts the power supply sources for each time period within the coordinated dispatch strategy. When PV power generation exceeds a preset threshold, it prioritizes allocating PV power to the operation of chiller units or the charging of cold storage devices, reducing the amount of electricity purchased from the grid. When PV power generation falls below the preset threshold, it switches to grid power supply, and determines the operating mode of the chiller units and cold storage devices based on the current electricity price signal.

[0055] Simultaneously, it receives regulation commands from the grid-side virtual power plant platform or demand response center. When a load reduction command is received, it utilizes the stored cooling capacity in the cold storage device to replace the immediate cooling output of the chiller unit, reducing the real-time power drawn by the system from the grid and providing peak-shaving regulation for the grid. When a load increase command is received, it increases the operating power of the chiller unit or starts the cold storage device to charge the cooling capacity, providing valley-filling regulation for the grid.

[0056] By integrating grid electricity prices, photovoltaic power generation, energy storage battery status of charge, and demand response commands, the system dynamically switches and combines various energy supply modes, including direct grid power supply, direct photovoltaic power supply, energy storage power release, and cold storage power release, and outputs comprehensive dispatch commands to each device for execution.

[0057] It should be noted that the aforementioned demand response interaction simultaneously optimizes revenue during execution. When participating in grid demand response, the sum of electricity cost savings and demand response compensation revenue is used as the optimization objective to determine the amount of power regulation and the time window for participation. The upper limit of the power regulation amount is jointly limited by the remaining cold storage capacity of the cold storage device and the safety constraints of the cold storage temperature, ensuring that the safety of the cold storage temperature is not affected while responding to grid regulation.

[0058] In this embodiment, to further reduce overall energy costs, the generation of integrated dispatch instructions also considers the charging and discharging scheduling of the energy storage battery system. During off-peak electricity prices and when photovoltaic power generation is insufficient, the energy storage battery is charged from the grid. During peak electricity prices, the energy storage battery releases power to the chiller unit, operating synchronously with the cooling energy storage device, further reducing the amount of electricity purchased from the grid during peak hours. The charging and discharging power of the energy storage battery is uniformly solved within the action space of the safety constraint deep reinforcement learning framework in step S302.

[0059] Step S500: Based on the execution result data of the collaborative scheduling strategy, perform cost accounting and strategy feedback optimization, and output a cost analysis report and updated scheduling parameters.

[0060] Collect data on the actual execution results of collaborative scheduling strategies and integrated scheduling instructions, including actual power consumption, actual electricity expenditure, actual charging and discharging capacity of cold storage devices, actual photovoltaic power consumption, and actual demand response compensation revenue for each time period.

[0061] The activity-based costing (ABC) approach breaks down the total operating cost within the scheduling cycle into sub-items such as energy cost, equipment depreciation cost, operation and maintenance cost, and carbon emission cost. Energy cost is calculated based on the actual electricity consumption and corresponding electricity price for each time period. Carbon emission cost is calculated based on the product of the grid carbon emission factor and the actual electricity consumption of the system for each time period. All of the above costs can be aggregated and traced by equipment type, time period, and cold storage zone, generating cost analysis reports.

[0062] It should be noted that in the above carbon emission accounting, the grid carbon emission factor for each time period is obtained from the regional grid emission data published by the grid side. The carbon emission cost is equal to the sum of the products of electricity consumption for each time period and the corresponding grid carbon emission factor for that time period, multiplied by the carbon price coefficient. The carbon emission accounting results generate a carbon emission report to support enterprises' carbon footprint management.

[0063] In this embodiment, to achieve continuous improvement of the scheduling strategy, based on the deviation between actual load and predicted load, and the deviation between actual electricity cost and estimated electricity cost in historical operating data, the input feature weights of each prediction model in step S200 and the reward function weight parameters of the security-constrained deep reinforcement learning framework in step S302 are updated on a rolling basis. Specifically, the actual operating data within the most recent preset period is appended to the training dataset, and the policy network in the security-constrained deep reinforcement learning framework is incrementally trained and updated to adapt the collaborative scheduling strategy to the long-term changing trends of electricity price structure, load pattern, and equipment status. At the same time, the expected electricity expenditure and total cost for future periods are predicted on a rolling basis, and the prediction results are compared with the cost estimate under the current scheduling parameter configuration to automatically recommend the adjusted scheduling parameter configuration.

[0064] Example 2

[0065] A cold chain storage center (hereinafter referred to as "Storage Center W") is equipped with 2 refrigeration units, 1 phase change energy storage device, a rooftop photovoltaic system, and energy storage battery packs. Storage Center W has a frozen storage area (target temperature range: -22°C to -18°C) and a refrigerated storage area (target temperature range: 0°C to 4°C). Storage Center W is connected to the peak-valley time-of-use electricity pricing system of its region and has been registered as a demand response resource on a virtual power plant platform. The scheduling cycle is set to 24 hours for a certain day in month A of 20XX, with a time interval of 1 hour. .

[0066] Implementation of step S100: Before the scheduling date, edge computing nodes at warehouse W continuously collect raw sensor data from various nodes, including refrigeration units, cold storage devices, and photovoltaic systems. Since the temperature sensor sampling period is 5 seconds, while the electricity price signal update period is 1 hour, the edge computing nodes use linear interpolation to align all signals to a 1-minute time resolution. Power sensor signals are filtered and denoised using a moving average with a window length of 5. The mean, variance, and rate of change are extracted from each 15-minute sampling window as feature vectors and uploaded to the cloud computing platform. Taking the period from 00:00 to 01:00 on the scheduling date as an example, some of the fused data received by the cloud platform is shown in the table below.

[0067] Table 1. Spatiotemporal tensor samples after fusion in step S100 (00:00 to 01:00 time period)

[0068] After Z-score standardization, the dimensional differences of each physical quantity are eliminated, and they are uniformly encoded into the input tensors of subsequent prediction models.

[0069] Implementation of step S200: The cloud computing platform uses a unified data spatiotemporal tensor to perform multi-timescale probability predictions for cooling load and electricity price.

[0070] In step S201, the Time-LLM large-scale time series model takes the daily goods inbound and outbound plan of the storage center W (an estimated batch of frozen goods, approximately 8.5 tons, will enter the warehouse between 14:00 and 17:00), outdoor weather forecast (afternoon high temperature 34℃), and historical load data for the same period as input, and outputs the probability distribution of cooling load for 24 time periods on a day-ahead scale. The upper and lower quantiles are extracted at a 90% confidence level to form the load prediction interval. The graph neural network uses refrigeration unit nodes, cold storage device nodes, and warehouse temperature nodes as vertices, and aggregates the features of adjacent nodes through graph convolutional layers, concatenating spatial correlation features to the Time-LLM prediction header to correct the prediction interval for the afternoon inbound disturbance period.

[0071] In step S202, the gradient boosting regression algorithm takes the historical electricity price sequence of the past 30 days and the date identifier (if the current day is a working day) as input, and outputs the expected electricity price value for each time period. Simultaneously, the cloud platform parses the demand response pre-notification issued by the virtual power plant platform, extracting the response time window as 18:00 to 20:00 and the response demand power as 150kW. Based on the historical triggering frequency, the probability of demand response triggering during this period is calculated to be 0.72.

[0072] In step S203, the load forecast confidence score is calculated using backtesting data of the actual load and forecast values ​​over the past 7 days, and the electricity price forecast confidence score is calculated using the actual electricity price and forecast values ​​over the past 7 days. Substituting these values ​​into the formula... ,in The total number of backtest samples, For the first The actual value for each time period This corresponds to the predicted value. Take... (7 days × 24 hours) Calculate the load forecast confidence score Electricity price forecast confidence score .

[0073] Table 2 shows the forecast output results for representative periods at the previous timescale.

[0074] The load forecast range for the 15:00 period reached 41.5kW, which is significantly wider than that during off-peak hours. This indicates that there is considerable uncertainty caused by inbound disturbances, and subsequent dispatching strategies will reserve more cooling capacity during this period.

[0075] Implementation of step S300: In step S301, the scheduling cycle Time step The safe temperature range for the freezer is -22°C to -18°C per hour, and the safe temperature range for the refrigerated warehouse is 0°C to 4°C per hour. Decision variables include the operating frequency of the two refrigeration units per time period (20Hz to 60Hz) and the charging and discharging power of the cold storage energy storage device (charge capacity up to 120kW, release capacity up to 100kW).

[0076] In step S302, the policy network of the security-constrained deep reinforcement learning framework receives a state vector consisting of the current cold storage capacity (312.4 kWh), the freezer temperature (-20.3℃), the current electricity price (0.38 yuan / kWh), the load forecast interval, and the electricity price forecast interval. Because... , Take the higher value, 0.60; because , Take 0.25, Let's set it to 0.15. The reward function is... ,in This represents the electricity cost for the current period. For the load variance, For the number of unit start-ups and shutdowns, all three items were standardized using Z-score before being substituted. The safety constraint layer detected that the cold storage capacity was close to the lower limit (predicted remaining 58kWh) between 19:00 and 20:00, and automatically increased the Lagrange penalty coefficient, driving the strategy to supplement the cold storage capacity between 16:00 and 18:00 in advance.

[0077] Table 3. Representative time period output of the collaborative scheduling strategy

[0078] In the charging and discharging power, "+" indicates charging and "-" indicates discharging. During the off-peak hours at 02:00, both units operate at high frequency and charge at full power. During the peak hours at 19:00, the units reduce to the lowest frequency and rely entirely on the cold storage device to release cold to maintain the temperature of the cold storage. The predicted temperature of the cold storage is -18.90℃, which is still within the safe range.

[0079] Implementation of step S400: At 14:30 on the dispatch day, the real-time power generation of the rooftop photovoltaic system reached 82kW, exceeding the preset threshold of 60kW. The cloud platform prioritized allocating this photovoltaic power to Unit 1, reducing grid power purchases by approximately 82kWh. At 19:05, the virtual power plant platform officially issued a load reduction command, requiring a load reduction of 150kW for 120 minutes from 18:00 to 20:00. The integrated dispatch command switched to a combined mode of "cold storage release + energy storage battery power release": the cold storage device released 100kW of cold, and the energy storage battery supplied power to the unit at 50kW. The grid power draw from the system decreased from the originally planned 182kW to 32kW, meeting the reduction requirements.

[0080] Table 4. Step S400 Integrated Dispatch Instructions (18:00 to 20:00 Demand Response Period)

[0081] During the demand response period, the power purchased by the grid was less than 35kW, which is about 82% lower than the current planned value, meeting the reduction requirements of the virtual power plant platform.

[0082] Implementation of step S500: After the scheduling day ends, the cloud platform collects the actual execution data for the entire day and uses the activity-based costing method for cost accounting. Energy costs are calculated based on the actual electricity consumption and corresponding electricity price for each time period, while carbon emission costs are calculated by multiplying the electricity consumption for each time period by the regional power grid carbon emission factor (average 0.581 kgCO2 / kWh) and then multiplying by the carbon price coefficient (55 yuan / ton).

[0083] Table 5 Summary of Daily Dispatch Cost Analysis Report

[0084] Demand response compensation revenue of RMB 187.50 was directly deducted from the overall net cost. Based on the deviation between today's actual load and the predicted load (actual load at 15:00 was 191.2kW, and the upper limit of the prediction was 198.3kW, with a deviation of approximately 3.6%), the cloud platform appended today's data to the training set to incrementally train the policy network. Updated to 0.88, and automatically recommends increasing the upper limit of charging and cooling power from 02:00 to 04:00 to 115kW on the next scheduling day to further utilize off-peak electricity price differences.

[0085] Throughout the implementation process, data originates from the raw sensor data acquired in step S100, and is then time-aligned and standardized to generate a unified spatiotemporal tensor. Step S200 outputs load forecast intervals and electricity price forecast intervals with confidence scores based on this tensor, quantifying and transmitting forecast uncertainty. Step S300 uses the forecast output and the current state of the cold storage device as inputs, and solves the collaborative scheduling strategy through a safety-constrained deep reinforcement learning framework, dynamically adjusting the weight coefficients of the confidence scores. , , This ensures that the scheduling strategy achieves an adaptive balance between economy and security. Step S400 integrates the collaborative scheduling strategy with real-time photovoltaic output and demand response instructions, outputs a comprehensive scheduling instruction, and executes it. Step S500 performs cost accounting on the execution results and feeds back the deviation data to the prediction model and strategy network, forming a complete closed-loop data flow from data acquisition, probability prediction, optimization decision-making, collaborative execution to cost assessment.

[0086] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0087] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0088] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart energy consumption scheduling and cost control system for cold chain refrigeration equipment, characterized in that, include: The data acquisition and fusion module is used to collect operating data of refrigeration units and cold storage energy storage devices in the cold chain refrigeration system, as well as time-of-use electricity price data. After time alignment and standardization processing, a unified data spatiotemporal tensor is generated. The probability prediction module is used to perform probability predictions on cooling load demand and electricity price at multiple time scales based on the unified data spatiotemporal tensor, output the load prediction interval and electricity price prediction interval for each time period, and calculate the confidence score of each prediction result. The collaborative scheduling optimization module is used to take the load prediction interval, the electricity price prediction interval, the confidence score and the current cold storage capacity of the cold storage energy storage device as inputs, with the minimization of total electricity cost as the primary optimization objective, and to perform multi-objective optimization under the hard constraint that the cold storage temperature does not exceed the safety threshold. It outputs the operating frequency of the refrigeration unit and the charging power or releasing power of the cold storage energy storage device in each time period, thus forming a collaborative scheduling strategy. The integrated scheduling execution module is used to dynamically switch between multiple energy supply modes based on the aforementioned collaborative scheduling strategy, combined with photovoltaic power generation and grid demand response instructions, output integrated scheduling instructions, and issue them for execution.

2. The intelligent energy consumption scheduling and cost control system for cold chain refrigeration equipment according to claim 1, characterized in that, The system collects operational data and time-of-use electricity price data from refrigeration units and cold storage devices in the cold chain refrigeration system. After time alignment and standardization, it generates a unified spatiotemporal tensor, which also includes: On the edge computing node, the raw sensor data is subjected to moving average filtering noise reduction processing, and the length of the filtering window is configured according to the noise characteristics of each sensor signal. The mean, variance and rate of change in each sampling window are extracted from the raw time series data as feature vectors, and the feature vectors are used to replace the raw high-frequency data and uploaded to the cloud computing platform. The time alignment process employs a linear interpolation algorithm to uniformly interpolate sensor data with different sampling frequencies to the same time resolution. Z-score standardization is performed on various types of sensor data to eliminate the dimensional differences between different physical quantities, and then the data is organized into the unified data spatiotemporal tensor according to the time dimension and the spatial dimension of each sensor node.

3. The intelligent energy consumption scheduling and cost control system for cold chain refrigeration equipment according to claim 1, characterized in that, Probabilistic forecasting of cooling load demand across multiple time scales, including: Cold chain storage data, cargo entry and exit plan data, historical outdoor temperature and humidity data, and weather forecast information are input into a large time series model. The large time series model performs feature encoding and sequence inference on the input multivariate time series data, and outputs the cold load probability distribution for each time period at three time granularities: day-ahead scale, intraday scale, and real-time scale. The upper quantile and lower quantile corresponding to the preset confidence level are extracted from the cold load probability distribution for each time period to form the load prediction interval for each time period. The feature vectors of each sensor node in the unified data spatiotemporal tensor are input into a graph neural network. The graph neural network forms a graph structure with each sensor node as a vertex and the physical connection relationship between nodes as edges. The features of adjacent nodes are aggregated through graph convolutional layers to output the feature representations of each node that fuse spatial correlation information. The feature representations of each node are then concatenated with the time series features output by the time series large model and re-input into the prediction head of the time series large model to output the updated probability distribution of cooling load for each time period.

4. The intelligent energy consumption scheduling and cost control system for cold chain refrigeration equipment according to claim 1, characterized in that, Probabilistic forecasting of electricity prices includes: Using historical time-of-use electricity price data and power grid dispatch announcements as input, a gradient boosting regression algorithm is employed to predict electricity prices for each future time period. The inputs to the gradient boosting regression algorithm are historical electricity price sequences, date type identifiers, and key fields from power grid dispatch announcements, and the output is the expected electricity price value for each future time period. Based on the prediction residual distribution of the gradient boosting regression algorithm, the upper and lower bounds of the electricity price prediction for each time period are calculated, forming the electricity price prediction interval. The system analyzes demand response event notifications issued by the grid side, extracts the response time window and response demand power parameters, and calculates the trigger probability of demand response events in future time periods based on the trigger frequency of historical demand response events and the current grid load status.

5. The intelligent energy consumption scheduling and cost control system for cold chain refrigeration equipment according to claim 1, characterized in that, The confidence score is calculated as follows: Based on the backtesting error statistics of each prediction model on recent historical data, a confidence score is calculated. The confidence score is equal to 1 minus the average absolute value of the relative error between the predicted value and the actual value in each backtesting period within a recent preset time window. Wherein, when the actual value of a certain backtesting period is zero, that period is not included in the error statistics. The confidence score ranges from 0 to 1. The larger the value, the higher the confidence level of the prediction. The confidence score is used as the weight coefficient of each prediction input in the subsequent multi-objective optimization solution.

6. The intelligent energy consumption scheduling and cost control system for cold chain refrigeration equipment according to claim 1, characterized in that, The multi-objective optimization solution employs a security-constrained deep reinforcement learning framework, which consists of a policy network, a value network, and a security constraint layer, wherein: The input to the strategy network is a state vector consisting of the current cold storage capacity, cold storage temperature, current electricity price, load prediction range, and electricity price prediction range, and the output is an action vector consisting of the operating frequency setpoint of each refrigeration unit and the charging and discharging power setpoint of the cold storage energy storage device. The input to the value network is the same as that of the policy network, and the output is the estimated cumulative reward for performing the corresponding action in the current state. The safety constraint layer performs temperature constraint verification on the action vector output by the policy network. The Lagrange relaxation method is used to embed the cold storage temperature safety constraint into the policy update process. When the predicted value of the cold storage temperature approaches the boundary of the cold storage temperature safety threshold, the penalty coefficient of the constraint term is automatically increased, driving the policy away from the constraint violation area.

7. The intelligent energy consumption scheduling and cost control system for cold chain refrigeration equipment according to claim 1, characterized in that, The reward function of the safety-constrained deep reinforcement learning framework consists of a weighted sum of the total electricity cost, the load power variance, and the compressor start-stop penalty. The three terms are Z-score standardized before being substituted into the reward function to eliminate the difference in dimensions. Among them, the weight coefficient of the total electricity cost item is positively correlated with the confidence score of the electricity price forecast, and the weight coefficients of the load power variance item and the compressor start-stop penalty item are positively correlated with the confidence score of the load forecast. When the confidence score is high, the corresponding weight coefficient of the forecast input increases, and the collaborative scheduling strategy uses the forecast results to perform peak-valley arbitrage. When the confidence score is low, the corresponding weight coefficient decreases, and the collaborative scheduling strategy prioritizes ensuring the safety constraints of the cold storage temperature.

8. The intelligent energy consumption scheduling and cost control system for cold chain refrigeration equipment according to claim 1, characterized in that, The collaborative scheduling strategy is decomposed and executed at three levels: day-ahead, intraday, and real-time. At the day-ahead level, based on the day-ahead load forecast interval and electricity price forecast interval, a macro power allocation scheme for each period of the next 24 hours is generated to determine the target operating status of chiller units and cold storage energy storage devices for each period. At the intraday level, every preset rolling cycle, the macroscopic power allocation scheme is rolled over and corrected based on the intraday scale prediction update data and actual operating deviations. At the real-time level, model predictive control is adopted, taking real-time sensor data and real-time scale prediction data as inputs, and outputting fine-tuning values ​​of control quantities within a short future time window to fine-tune the operating frequency of the chiller unit and the charging and discharging power of the cold storage energy storage device in real time.

9. The intelligent energy consumption scheduling and cost control system for cold chain refrigeration equipment according to claim 1, characterized in that, Dynamically switch between multiple energy supply modes, including: When the photovoltaic power generation is higher than the preset threshold, the photovoltaic power is allocated to the operation of the chiller unit or the cooling energy storage device; when the photovoltaic power generation is lower than the preset threshold, the power is switched to grid power and the operating mode is determined according to the current electricity price signal; during the peak electricity price period, the energy storage battery releases electricity to the chiller unit and operates synchronously with the cooling energy storage device. When a load reduction command is received from the grid, the stored cold energy in the cold storage device is used to replace the immediate cooling output of the refrigeration unit; when a load increase command is received, the operating power of the refrigeration unit is increased or the cold storage device is started to charge the cold energy. The output of the integrated scheduling instruction also includes: collecting actual execution result data, adding the actual operation data within the most recent preset period to the training dataset, and incrementally training and updating the strategy network in the multi-objective optimization solution so that the collaborative scheduling strategy can adapt to the long-term changing trends of electricity price structure and load pattern.

Images