Multi-device power consumption monitoring, prediction and control system and method

Abstract

This invention discloses a multi-device power consumption monitoring, prediction, and control system and method, belonging to the field of intelligent manufacturing control. The system includes an information perception layer, a business collaboration layer, and an edge intelligence layer. It collects voltage, current, and power data through smart sockets deployed at the front end of the power supply circuit of production equipment. Combined with work order information provided by the MES system, the edge computing server performs work order energy consumption calculation, multi-step power prediction, and control decisions. The method includes: real-time acquisition of equipment power data, calculation of unit product energy consumption and electricity cost, prediction of multi-step power values ​​based on equipment type-specific model weights, and execution of differentiated control strategies based on the predicted total power. This invention achieves refined calculation of work order-level energy consumption and electricity cost, improves the accuracy of multi-device power consumption prediction and forward-looking control capabilities, and effectively supports load peak shaving and energy efficiency optimization.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing control, specifically relating to a multi-device power consumption monitoring, prediction and control system and method. Background Technology

[0002] In the field of intelligent manufacturing, monitoring and managing the energy consumption of production equipment has always been a key aspect for enterprises to achieve refined operations and energy conservation. Traditional methods mainly rely on manual inspections and periodic meter readings, using scattered meter readings combined with experience to judge energy consumption. This approach not only has long data collection cycles and poor real-time performance, but also makes it difficult to accurately assess the overall energy efficiency of multiple production equipment operating collaboratively. As production lines expand and equipment types increase, relying solely on manpower can no longer meet the needs for timely response and in-depth analysis of dynamic changes in energy consumption.

[0003] Therefore, industrial energy management systems are gradually being introduced into practical applications. These systems typically collect key power consumption parameters based on sensor networks and use data platforms for preliminary aggregation and display, enabling centralized monitoring of energy consumption in individual devices or areas. Some systems also incorporate basic data analysis modules, capable of trend analysis of historical energy consumption data to assist managers in identifying abnormal energy consumption phenomena. The deployment of such systems has, to some extent, improved the visibility and management efficiency of energy consumption data, providing preliminary data support for energy-saving decisions.

[0004] However, existing industrial energy management systems still have significant shortcomings when dealing with complex operating scenarios involving multiple devices. Firstly, at the energy consumption accounting level, existing solutions struggle to accurately link total energy consumption to specific production work orders and the output of qualified products, resulting in coarse-grained calculations of unit product energy consumption and electricity costs, lacking direct guidance for production decisions. Secondly, in terms of forecasting capabilities, existing solutions lack forecasting methods that can take into account the unique operating characteristics of various devices and efficiently capture their long-term time-series dependencies, leading to large power consumption prediction errors and insufficient generalization ability. Consequently, they also lack closed-loop control capabilities based on power consumption prediction, failing to proactively and differentiatedly intervene in non-critical equipment or issue timely warnings before the total load is about to exceed limits, thus failing to achieve reliable peak shaving and valley filling without affecting core production. Summary of the Invention

[0005] This invention proposes a multi-device power consumption monitoring, prediction, and control system and method. Its purpose is to: calculate the unit product energy consumption and electricity cost associated with specific production work orders and qualified output; improve the prediction accuracy and generalization ability of power consumption of different types of production equipment, effectively capture their long-term time-series dependencies, and establish a forward-looking closed-loop control mechanism based on power consumption prediction, so as to achieve differentiated equipment intervention or alarm before the total load exceeds the limit, thereby achieving reliable load shaving and energy efficiency optimization while ensuring the continuity of core production.

[0006] The technical solution of this invention is as follows:

[0007] A multi-device power consumption monitoring, prediction, and control system includes:

[0008] The information sensing layer includes smart sockets deployed at the front end of the power supply circuits of each production device, which are used to collect and upload the voltage, current and power data of the corresponding production device to the edge computing server, as well as to receive and execute control commands from the edge computing server.

[0009] The business collaboration layer includes the Manufacturing Execution System (MES), which provides work order information to the edge computing server. The work order information includes the start and end times of the work order execution period, the list of associated production equipment, and the qualified output.

[0010] Edge intelligence layer, including edge computing servers;

[0011] The edge computing server includes:

[0012] The work order energy consumption calculation module is used to calculate the unit product energy consumption and unit product electricity cost based on the data collected by the smart socket and the work order information provided by the MES system.

[0013] The power prediction module is used to load the corresponding exclusive model weights into the power prediction model according to the type of production equipment, and use the power prediction model to predict the multi-step power prediction values ​​of each production equipment within a predetermined time period based on historical power sequences and time context features.

[0014] The control decision module is used to accumulate the multi-step power prediction values ​​of all production equipment to be operated to obtain the predicted total power curve, and when the predicted total power exceeds the preset power demand threshold, it executes differentiated control strategies according to the control attributes of each production equipment.

[0015] As a further improvement to the multi-device power consumption monitoring, prediction, and control system, the processing procedure of the work order energy consumption calculation module includes the following steps:

[0016] Step A1: The work order energy consumption calculation module obtains the information of the currently executing work orders in real time, and combines it with the real-time status data of the production equipment to filter out the set of target production equipment that is in the "running" state during the work order's execution period;

[0017] Step A2: For each piece of production equipment in the target production equipment set, extract its active power time-series data segment during the work order's execution period, and calculate the actual energy consumption value of the production equipment in that work order:

[0018] ;

[0019] In the above formula, Indicates the first The actual energy consumption of the production equipment during the work order's operating period; Indicates the first The equipment in the first The active power value at each sampling time; Indicates the sampling time interval; This indicates the total number of sampling points for the production equipment during the work order's running period;

[0020] Step A3: Select all of the target production equipment set The total energy consumption of this work order is obtained by summing the actual energy consumption values ​​of each production device:

[0021] ;

[0022] In the above formula, This indicates the total energy consumption of the work order;

[0023] Step A4: Based on the start and end times of the work order's runtime, match the time-of-use pricing strategy and segment the work order's runtime into time-based segments. There are several electricity price ranges, and the cumulative energy consumption of the target production equipment set within each range is calculated. The cumulative energy consumption for each electricity price range is recorded as follows: ;

[0024] Step A5: Based on the start and end times of each electricity price range, obtain the corresponding unit electricity price. The unit electricity price for each electricity price range is denoted as: ;

[0025] Step A6: Accumulate energy consumption for each time period The corresponding unit electricity price After multiplying and adding them together, we get the total electricity cost for this work order:

[0026] ;

[0027] In the above formula, This indicates the total electricity cost for the work order.

[0028] Step A7: Read the qualified output corresponding to this work order from the MES system, and calculate the total energy consumption of the work order. Divide by the qualified output to get the unit product energy consumption, and then calculate the total electricity cost. Divide by the qualified output to get the unit product electricity cost.

[0029] As a further improvement to the multi-device power consumption monitoring, prediction and control system, the power prediction model includes two dilated causal convolution-temporal channel attention blocks, a gated recurrent unit, a temporal attention enhancement unit and a fully connected layer connected in sequence.

[0030] As a further improvement to the multi-device power consumption monitoring, prediction and control system: the input features of the dilated causal convolution-temporal channel attention block are obtained by convolution to obtain the first feature, and by two cascaded dilated causal convolution units to obtain the second feature. Then the first feature and the second feature are added together to obtain the output features of the dilated causal convolution-temporal channel attention block.

[0031] Each dilated causal convolutional unit contains a dilated causal convolutional layer, a weight normalization layer, a ReLU activation function, a dropout layer, and a temporal-channel attention module connected in sequence.

[0032] As a further improvement to the multi-device power consumption monitoring, prediction and control system: the time-channel attention module includes two parallel branches, a channel attention branch and a time-series attention branch. The input features of the time-channel attention module are input into the two branches respectively. The features output by the channel attention branch are multiplied element-wise with the features output by the time-series attention branch. The result of the multiplication is then passed through the Sigmoid activation function to obtain a third feature. This third feature is then multiplied by the input features of the time-channel attention module to obtain the output features of the time-channel attention module.

[0033] As a further improvement to the multi-device power consumption monitoring, prediction and control system, the power prediction module also has a model hot update mechanism. When the power prediction error of the production equipment continuously exceeds the preset tolerance threshold, the model update process is automatically triggered to fine-tune the exclusive model weights of the same type of production equipment and dynamically load new exclusive model weights through uninterrupted hot replacement technology.

[0034] As a further improvement to the multi-device power consumption monitoring, prediction, and control system, the differentiated control strategies executed by the control decision module include:

[0035] For production equipment with the control attribute set to "automatic control enabled", generate a power-off or delayed start command and send it to the corresponding smart socket for execution;

[0036] For production equipment with control attributes of "critical equipment" or "automatic intervention prohibited", structured early warning information is generated and pushed to the terminal of the preset responsible personnel.

[0037] As a further improvement to the multi-device power consumption monitoring, prediction and control system, it also includes a user interaction layer for dynamically presenting the processing results of the edge computing server in a graphical interface.

[0038] A method for monitoring, predicting, and controlling the power consumption of multiple devices, based on the aforementioned system for monitoring, predicting, and controlling the power consumption of multiple devices; the method specifically includes:

[0039] Step S1: Through smart sockets deployed at the front end of the power supply circuit of each production equipment, the voltage, current and power data of the production equipment are collected in real time and uploaded to the edge computing server;

[0040] Step S2: The work order energy consumption calculation module calculates the unit product energy consumption and unit product electricity cost based on the information from the smart socket and the MES system;

[0041] Step S3: The power prediction module predicts the multi-step power prediction values ​​of each production device based on the dedicated model weights;

[0042] Step S4: The control decision module obtains the predicted total power based on the multi-step power prediction value, compares the predicted total power with the power demand threshold, obtains the control strategy, and executes it.

[0043] As a further improvement to the multi-device power consumption monitoring, prediction, and control method, it also includes:

[0044] The system analyzes the operating status data of each production device in real time and identifies production devices that have been in standby mode for a long time but have no production tasks associated with them. If the standby time of a production device exceeds a preset threshold, it performs differentiated processing based on its control attributes: for non-critical devices with the control attribute of "allow automatic control", it automatically issues a power-off command; for production devices with the control attributes of "critical device" or "prohibit automatic intervention", it generates a standby timeout warning and pushes it to the terminal of the relevant responsible personnel.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. This invention collects real-time active power data at the device level through smart sockets and combines it with work order information provided by the Manufacturing Execution System (MES). This allows for the precise identification of production equipment in an "operating" state during the effective execution period of the work order, along with its corresponding power timing data. The resulting data allows for the calculation of unit product energy consumption and unit product electricity cost. This method achieves precise traceability of energy consumption and cost at the work order level, providing a reliable basis for product-level energy efficiency assessment and cost allocation.

[0047] 2. This invention proposes a mechanism for loading device-specific model weights under a unified deep learning architecture. The power prediction model integrates dilated causal convolution, a time-channel dual-dimensional attention module, and a gated recurrent unit. It can expand the receptive field through dilated convolution to capture long-term temporal dependencies, and dynamically identify key historical time steps and feature channels with the help of the attention mechanism, thereby effectively adapting to the load characteristics of different types of devices.

[0048] 3. This invention achieves dedicated model weight management for different types of production equipment within a unified framework, balancing prediction accuracy, cost, and ease of deployment and maintenance, and improving the model's generalization ability.

[0049] 4. Based on the multi-step power prediction results, the predicted total power curve is generated by accumulating the predicted power of all equipment to be operated. When the predicted total power exceeds the power demand threshold, differentiated regulation is performed according to the preset control attributes of the equipment. This realizes proactive intervention and manual coordination before the load exceeds the limit, effectively achieving peak shaving and valley filling while ensuring the continuity of core production, and reducing energy costs and operational risks.

[0050] 5. The present invention also has the ability to update the model hot. By continuously monitoring the prediction error and automatically triggering incremental training and uninterrupted hot replacement when the error exceeds the tolerance threshold, the power prediction model can adapt to changes in the operating status of the equipment, maintain long-term prediction accuracy, and improve the robustness and adaptability of the system in dynamic industrial environments.

[0051] 6. This invention also expands the intelligent monitoring and intervention capabilities for equipment standby status. By identifying equipment that has been in standby for a long time and has no production tasks associated with it, and by executing automatic power-off or sending early warnings based on its control attributes, it effectively reduces hidden energy consumption and further improves the closed loop of energy management from operation to standby. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the architecture of a multi-device power consumption monitoring, prediction and control system in a specific implementation.

[0053] Figure 2 This is a schematic diagram of the power prediction model architecture;

[0054] Figure 3 A schematic diagram of the architecture of dilated causal convolution-temporal channel attention block;

[0055] Figure 4 This is a schematic diagram of the architecture of the time-channel attention module. Detailed Implementation

[0056] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0057] like Figure 1 As shown, a multi-device power consumption monitoring, prediction and control system adopts a layered architecture design, including an information perception layer, a business collaboration layer, an edge intelligence layer, a data support layer and a user interaction layer, and constructs a closed-loop management mechanism covering the entire process of "perception-accounting-prediction-control-traceability".

[0058] (1) Information perception layer.

[0059] The information sensing layer includes smart sockets deployed at the front end of the power supply circuits of each production device. These smart sockets have power data acquisition, wireless data transmission, and remote control functions. They are used to collect and upload voltage, current, and power data of the corresponding production device to the edge computing server, and also to receive remote control commands from the edge computing server to control the power supply circuit in which they are located.

[0060] Specifically, the smart socket incorporates a high-precision power metering chip (BL0942 in this embodiment), which can collect real-time voltage, current, and active power data of connected devices and upload structured data to the edge computing server of the edge intelligence layer via a Wi-Fi network based on the MQTT IoT protocol. Simultaneously, the smart socket integrates a microcontroller, a communication unit, and a relay control unit. The local microcontroller can perform preliminary calibration and filtering preprocessing on the collected data (such as removing instantaneous noise and normalizing dimensions). The communication unit and relay control unit support remotely receiving commands to execute power-off or delayed start operations and provide local operating status feedback through status indicator lights, achieving integrated measurement and control.

[0061] (2) Business collaboration layer.

[0062] The business collaboration layer includes the MES (Manufacturing Execution System), which serves as the central hub for production information. It provides work order information to the edge computing server, including the start and end times of the work order's execution period, a list of associated production equipment, qualified output, and the operating status of each production equipment within the work order cycle (e.g., "running," "standby," or "stopped"). This structured data provides the necessary production semantic support for subsequent accurate energy consumption calculation and power prediction.

[0063] (3) Edge intelligence layer.

[0064] The edge intelligence layer is the core decision-making unit, comprising several edge computing servers. Each edge computing server embeds three main functional modules: a work order energy consumption calculation module, a power prediction module, and a control decision module.

[0065] (3.1) Work order energy consumption calculation module.

[0066] The work order energy consumption calculation module is used to calculate the unit product energy consumption and unit product electricity cost based on information from smart sockets and the MES system. Specifically, the work order energy consumption calculation module obtains work order information from the MES system, filters out the effective power data segments of the production equipment associated with the work order in the "running" state, and uses the trapezoidal integral method to accurately calculate the actual energy consumption of each production equipment within the work order cycle based on the information collected by the smart socket; combined with the time-of-use electricity pricing strategy (which can be dynamically connected to an external power trading platform), it calculates the total electricity cost of the work order, and outputs the unit product energy consumption and unit product electricity cost based on the qualified output recorded by the MES system.

[0067] The work order energy consumption calculation module not only supports accurate measurement of the energy consumption of a single device, but more importantly, it enables the aggregation of energy consumption and refined calculation of electricity costs when multiple devices collaborate to complete the same production work order, thereby outputting unit product energy cost indicators. The specific processing includes the following steps:

[0068] Step A1: The work order energy consumption calculation module obtains the currently executing work order information from the MES system in real time via a RESTful API. This information includes the work order number, start and end times of the work order execution period, product model, qualified output, and a list of associated production equipment. Subsequently, combined with real-time status data of the production equipment (such as "running / standby / fault" signals reported by PLCs or sensors), the module filters out the set of target production equipment that is in the "running" state during the work order execution period, excluding equipment that is idling, debugging, or shut down, ensuring that subsequent energy consumption calculations only apply to effective production activities.

[0069] Step A2: For each piece of production equipment in the target production equipment set, extract its active power time-series data segment during the work order's runtime (sampling frequency can be from 1 second to 1 minute). Integrate the power curve using the trapezoidal numerical integration method to calculate the actual energy consumption of the production equipment in that work order. The calculation formula is as follows:

[0070] ;

[0071] In the above formula, Indicates the first The actual energy consumption of the production equipment during the work order's operating period; Indicates the first The equipment in the first The active power value at each sampling time; Indicates the sampling time interval; This indicates the total number of sampling points for the production equipment during the work order's running period.

[0072] Step A3: Select all of the target production equipment set The total energy consumption of this work order is obtained by summing the actual energy consumption values ​​of each production device:

[0073] ;

[0074] In the above formula, This indicates the total energy consumption of the work order.

[0075] Step A4: Based on the start and end times of the work order's runtime, automatically match the time-of-use pricing strategy configured by the local power grid or the enterprise (such as peak, flat, and valley time definitions), and segment the work order's runtime into time-based segments. There are several electricity price ranges, and the cumulative energy consumption of the target production equipment set within each range is calculated. The cumulative energy consumption for each electricity price range is recorded as follows: .

[0076] Step A5: Based on the start and end times of each electricity price range, dynamically obtain the corresponding unit electricity price from the built-in electricity price strategy table or an external power trading platform (such as the provincial power trading center API). The unit electricity price for each electricity price range is denoted as: Unit: Yuan / kWh. Supports multiple pricing models including fixed electricity price, tiered electricity price, and day-ahead market floating price.

[0077] Step A6: Accumulate energy consumption for each time period The corresponding unit electricity price After multiplying and adding them together, we get the total electricity cost for this work order:

[0078] ;

[0079] In the above formula, This indicates the total electricity cost for the work order.

[0080] Step A7: Read the qualified output corresponding to this work order from the MES system (excluding defective and reworked products), and calculate the total energy consumption of the work order. Divide by the qualified output to get the unit product energy consumption, and then calculate the total electricity cost. Divide the output by the qualified output to obtain the unit electricity cost (yuan / unit), which serves as the core indicator for measuring the energy efficiency and economy of the product.

[0081] Step A8: Save and display the calculation results. Specifically, the unit product energy consumption, unit product electricity cost, and detailed data from the calculation process are transmitted back to the MES system via a standardized interface, embedding them into its capacity accounting, cost center, or performance evaluation modules. Simultaneously, they are pushed to a visualization platform, displayed as work order-level energy consumption reports, equipment energy efficiency rankings, and unit cost trend charts. The platform supports filtering and querying by workshop, production line, and time period, and can export PDF or Excel reports for management to conduct energy-saving diagnostics and production scheduling optimization decisions. At the same time, abnormal cost data (such as deviations from the historical average by ±20%) are automatically marked, triggering a manual review process to ensure data reliability.

[0082] (3.2) Power prediction module.

[0083] The power prediction module is used to run a power prediction model, which can load specific model weights corresponding to different types of production equipment.

[0084] Specifically, when making predictions, the power prediction model loads corresponding exclusive model weights according to the type of production equipment to be predicted. The input of the power prediction model is the historical power sequence (such as data from the past 60 minutes) and time context features (such as the current hour, day of the week, whether it is a holiday, etc.). The output is the multi-step power prediction value divided into fixed time steps (such as one step every 5 minutes) within a predetermined time period in the future (such as the next 30 minutes).

[0085] like Figure 2 The power prediction model consists of two dilated causal convolutional-temporal channel attention blocks (TCN-TCA Block), a gated recurrent unit (GRU), a temporal attention enhancement unit, and a fully connected layer connected in sequence.

[0086] The input features of the power prediction model are first processed by two cascaded dilated causal convolutions and temporal channel attention blocks to extract features.

[0087] Furthermore, such as Figure 3The input features of the dilated causal convolution-temporal channel attention block are processed through two concatenated dilated causal convolutional units to obtain the first feature, and through two concatenated dilated causal convolutional units to obtain the second feature. The first and second features are then summed to obtain the output features of the dilated causal convolution-temporal channel attention block. Each dilated causal convolutional unit contains a dilated causal convolutional layer, a weight normalization layer, a ReLU activation function, a dropout layer, and a temporal-channel attention module connected in sequence. This design effectively expands the receptive field, captures long-term dependencies, strictly maintains temporal causality (i.e., does not rely on future information), and significantly improves training stability and generalization ability: the first dilated causal convolutional unit focuses on extracting local temporal patterns, while the second dilated causal convolutional unit focuses on dynamic features over longer periods.

[0088] Furthermore, such as Figure 4 The time-channel attention module comprises two parallel branches: a channel attention branch and a time-series attention branch. The input features of the time-channel attention module are fed into both branches. The channel attention branch dynamically evaluates the importance of each feature channel (such as voltage, current, and power) to the prediction task through global average pooling and a fully connected layer, adaptively enhancing its sensitivity to key signals such as load mutations or voltage fluctuations. The time-series attention branch, through a self-attention mechanism, calculates the correlation between historical time steps and the current prediction target, identifies the most influential time periods (such as periodic peaks or start-stop transients within the first 15 minutes), and dynamically increases their weights. The features output by the channel attention branch and the features output by the time-series attention branch are multiplied element-wise. The result of this multiplication is then processed through a sigmoid activation function to obtain a third feature. This third feature is then multiplied by the input features of the time-channel attention module to obtain the output features of the time-channel attention module.

[0089] like Figure 2 The features extracted through two cascaded dilated causal convolutions and temporal channel attention blocks are then input into a gated recurrent unit. Its gating mechanism adaptively adjusts the fusion ratio of historical states and current input information, further enhancing the modeling capability for complex operating modes. The features output from the gated recurrent unit are then attention-weighted by a temporal attention enhancement unit before being input into a fully connected layer, yielding the output of the power prediction model—the power prediction values ​​for multiple future time steps—providing a highly reliable decision-making basis for forward load regulation.

[0090] The deep learning model used in the power prediction module is based on a unified network architecture design, which supports dynamic loading of exclusive model weights according to device type, thereby achieving high-precision, multi-step device power prediction under the premise of shared structure.

[0091] During the training phase of the power prediction model, the system first collects historical power data from all equipment in the plant for one week and groups it according to equipment type. The raw data is normalized to eliminate dimensional differences, and then a sliding window strategy is used to segment the continuous time series into fixed-length sample segments. The dataset for each equipment type is divided into training, validation, and test sets in a 7:2:1 ratio. The input to the power prediction model includes historical power sequences and time context features (such as hour, weekday, and holiday identifiers). Through end-to-end training, dedicated model weights are independently optimized and generated for each equipment type under a unified architecture. After training, the dedicated model weights for each equipment type are stored as files on the edge computing server and can be accessed by the power prediction module as needed via a local API or lightweight HTTP interface. This "unified architecture, dedicated model weights" design ensures the model's adaptability to different equipment load characteristics while significantly reducing deployment and maintenance complexity, facilitating subsequent model updates and horizontal scaling.

[0092] It should be noted that compared to traditional time-series prediction models such as LSTM, standard TCN, or pure Transformer, the unified deep learning architecture adopted in this system has significant advantages in equipment power prediction tasks: While LSTM can model time-series dependencies, its serial computation efficiency is low, and it is difficult to accurately capture local abrupt changes such as equipment start-up and shutdown; while standard TCN has parallelism and a long receptive field, it lacks the ability to differentiate the modeling of multi-dimensional input features (such as voltage, current, and time context); and pure Transformer faces problems such as easy overfitting with small samples, high computational overhead, and difficulty in deployment in industrial edge scenarios. This system, through a hybrid structure of "TCN-TCA Block + GRU + lightweight attention", maintains temporal causality and efficient inference, while utilizing a time-series-channel dual-dimensional attention mechanism to adaptively focus on the historical time steps (such as periodic peaks) and key feature channels (such as load abrupt change signals) that have the greatest impact on the prediction results, thereby more accurately characterizing the complex power consumption patterns of different types of production equipment such as CNC machine tools, air compressors, and injection molding machines. Meanwhile, the design paradigm of "unified network architecture + device type-specific model weights" avoids the engineering complexity of developing separate models for each type of device, while ensuring high-precision adaptation to the unique load patterns of each device.

[0093] Furthermore, this embodiment also provides a hot update mechanism for the power prediction model: continuously calculating the mean absolute error (MAE) between the multi-step power prediction values ​​and the actual observed power values ​​of each production device; when the mean absolute error of a certain production device exceeds a preset tolerance threshold N times consecutively (e.g., N=5), the model update process is automatically triggered: notifying the responsible personnel to upload a new training dataset, or calling the edge incremental training module to obtain stored high-quality observation data as a training dataset, and fine-tuning the exclusive model weights for the same type of production device based on the training dataset; after training, a new version of the exclusive model weights is generated, and the new exclusive model weights are dynamically loaded through uninterrupted hot replacement technology to ensure the continuous availability of the prediction service. At the same time, the version number, update time, performance indicators, and triggering reasons of the exclusive model weights are written to a persistent database to support full lifecycle traceability and auditing. This mechanism enables the system to have long-term adaptive and self-evolving capabilities, ensuring the continuous stability of prediction accuracy in complex and dynamic manufacturing environments.

[0094] (3.3) Control Decision Module.

[0095] The control decision module is used to compare the predicted total power with the power demand threshold, obtain the control strategy, and execute it.

[0096] Specifically, the control decision module accumulates the multi-step power prediction values ​​of all production equipment to be operated to obtain the total predicted power curve, which includes the total predicted power for each future time period. If the total predicted power for a certain future time period exceeds the preset power demand threshold (such as the upper limit of transformer capacity, the maximum contract demand, the peak electricity price trigger threshold, etc.), a differentiated control strategy is formulated and executed according to the preset control attributes of the equipment. Otherwise, the current operating state is maintained, and only the predicted data is recorded for subsequent analysis.

[0097] The control strategy is formulated as follows: For production equipment with the control attribute of "allow automatic control", a power-off or delayed start command is generated and sent to the corresponding smart socket for execution, achieving millisecond-level response; For production equipment with the control attribute of "critical equipment" or "prohibit automatic intervention" (such as CNC machine tools and injection molding machines that are processing high-value products), a structured early warning information (including equipment name, predicted over-limit time, and suggested measures) is generated and pushed to the terminal of the preset responsible personnel (such as team leader, energy manager, or production scheduler), and manual intervention is used to evaluate and decide whether to adjust the production schedule or switch to standby production equipment, so as to achieve flexible closed-loop control.

[0098] (4) Data support layer.

[0099] The data support layer is a database storage module that communicates with the edge computing server. It adopts a layered collaborative storage architecture, including a high-speed cache layer and a persistent storage layer. The high-speed cache layer uses a Redis database to cache the original electrical parameter data of the devices in real time to meet the low-latency data access requirements of the power prediction module. The persistent storage layer uses a relational database, specifically PostgreSQL in this embodiment, to archive energy consumption calculation results, device metadata, model version information, and control logs by work order dimension. The system is also configured with a timed scheduling service to listen for the "work order completed" event published by the MES system. Once this event is detected, it automatically triggers the batch transfer of relevant data from the cache layer to the persistent layer and releases cache space, improving resource utilization efficiency.

[0100] (5) User interaction layer.

[0101] The user interaction layer is a visualization platform that interfaces with the edge computing server. It dynamically presents the processing results of the edge computing server through a graphical interface, including but not limited to key information such as real-time power curves of devices, work order-level energy consumption and cost analysis, future load forecast trends, and control command execution logs. The visualization platform supports historical tracing, energy efficiency benchmarking, and cost allocation comparisons by product, production line, shift, and time period, providing data support and decision-making basis for production scheduling, energy management, and process optimization.

[0102] This invention provides a method for monitoring, predicting, and controlling the power consumption of multiple devices. Its core process revolves around a closed-loop management mechanism encompassing "sensing—calculation—prediction—control—traceability." This method is implemented based on the aforementioned system architecture and specifically includes the following steps:

[0103] Step S1: Real-time sensing and uploading of device-level power data.

[0104] By deploying smart sockets at the front end of the power supply circuits of each production device, the voltage, current, and active power data of the corresponding production device are collected in real time (at a frequency of 1Hz or higher). The smart sockets then upload the collected structured power data to the edge computing server of the edge intelligence layer via a Wi-Fi network based on the MQTT IoT protocol.

[0105] Step S2: The work order energy consumption calculation module calculates the unit product energy consumption and unit product electricity cost based on the information from the smart socket and the MES system.

[0106] The edge computing server obtains real-time production work order information from the MES system via a RESTful API, including the start and end times of the work order's execution period, the list of associated production equipment, and the qualified output. Based on the real-time operating status of the production equipment, it filters out the set of target production equipment in the "running" state within the work order's execution period and their corresponding effective power time-series data segments. Then, it calculates the actual energy consumption of the target production equipment set within the work order's execution period, uses time-of-use pricing strategies to calculate the total electricity cost of the work order, and outputs the unit product energy consumption and unit product electricity cost based on the qualified output.

[0107] Step S3: The power prediction module predicts the multi-step power prediction values ​​of each production device based on the exclusive model weights.

[0108] For production equipment that needs to participate in operation scheduling, the edge computing server calls the power prediction module, loads the corresponding exclusive model weights into the power prediction model according to the type of production equipment to be predicted, and then inputs the historical power sequence and time context features of the production equipment into the power prediction model, and outputs multi-step power prediction values ​​divided by fixed time steps within a future predetermined time period.

[0109] Step S4: The control decision module obtains the predicted total power based on the multi-step power prediction value, compares the predicted total power with the power demand threshold, obtains the control strategy, and executes it.

[0110] The control decision module accumulates the multi-step power prediction values ​​of all production equipment to be operated to obtain the predicted total power curve. If the predicted total power for a future period exceeds the preset power demand threshold, a differentiated control strategy is executed according to the preset control attributes of each production equipment: for production equipment with the control attribute of "allow automatic control", control commands are generated and sent to its smart socket; for production equipment with the control attribute of "critical equipment" or "prohibit automatic intervention", structured early warning information is generated and pushed to the terminal of the preset responsible personnel.

[0111] Step S5: Visualize and interactively trace multi-dimensional data through a visualization platform.

[0112] The energy consumption calculation results, multi-step power prediction values, control command logs, and early warning information generated in steps S3 to S5 are uniformly pushed to the user interaction layer's visualization display platform. This platform dynamically presents the above information in a graphical interface, supporting historical tracing, energy efficiency benchmarking, and cost allocation comparisons by product, production line, shift, time period, and other dimensions.

[0113] The multi-device power consumption monitoring, prediction, and control method also includes:

[0114] Continuous monitoring of production equipment in standby mode and intelligent intervention based on equipment attributes and preset strategies are implemented. Specifically, the operating status data of each production device is analyzed in real time to identify production devices that have been in "standby" for extended periods without any associated production tasks. If the standby time of a production device exceeds a preset threshold (e.g., 30 minutes), differentiated processing is performed based on its control attributes: for non-critical equipment with the control attribute of "automatic control allowed," a power-off command is automatically issued to eliminate invalid standby energy consumption; for "critical equipment" or equipment with "automatic intervention prohibited," a standby timeout warning is generated and pushed to the terminals of relevant responsible personnel, prompting them to confirm whether shutdown or maintenance needs to be arranged. This mechanism effectively avoids "hidden energy consumption" waste, further improves the overall energy efficiency of the factory, and realizes a refined energy management closed loop from "operation" to "standby / shutdown" throughout the entire lifecycle.

[0115] It should be noted that, as will be apparent to those skilled in the art, the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics thereof. The scope of the present invention is defined by the claims rather than the foregoing description.

Claims

1. A multi-device power consumption monitoring, prediction and regulation system, characterized in that, The method comprises the following steps: An information perception layer comprises intelligent sockets deployed in front of power supply circuits of production equipment, which are used to collect and upload voltage, current and power data of the corresponding production equipment to an edge computing server, and receive and execute control instructions from the edge computing server; A business collaboration layer comprises a manufacturing execution system (MES) system, which is used to provide work order information to the edge computing server, the work order information comprising work order runtime period start and end time, associated production equipment list and qualified production quantity; An edge intelligence layer comprises an edge computing server; The edge computing server comprises: A work order energy consumption accounting module, which is used to calculate unit product energy consumption and unit product electricity cost according to the data collected by the intelligent sockets and the work order information provided by the MES system; A power prediction module, which is used to load corresponding exclusive model weights into a power prediction model according to production equipment types, and use the power prediction model to predict multi-step power prediction values of each production equipment in a future predetermined time period based on historical power sequences and time context features; The power prediction model comprises two dilated causal convolution-temporal channel attention blocks connected in sequence, a gated recurrent unit, a temporal attention enhancement unit and a fully connected layer; The input features of the dilated causal convolution-temporal channel attention block are added together after being obtained through convolution and two series-connected dilated causal convolution units to obtain the output features of the dilated causal convolution-temporal channel attention block; Each dilated causal convolution unit comprises dilated causal convolution layers, weight normalization layers, ReLU activation functions, dropout layers and temporal-channel attention modules connected in sequence; The temporal-channel attention module comprises two parallel branches, i.e., a channel attention branch and a temporal attention branch, and the input features of the temporal-channel attention module are input into the two branches; the features output by the channel attention branch and the features output by the temporal attention branch are multiplied element by element, and the multiplication result is then input into a Sigmoid activation function to obtain third features; the third features are multiplied with the input features of the temporal-channel attention module to obtain the output features of the temporal-channel attention module; The power prediction module also has a model hot update mechanism; when the power prediction error of a production equipment continuously exceeds a preset tolerance threshold, a model update process is automatically triggered to fine-tune the exclusive model weights of the same type of production equipment, and new exclusive model weights are dynamically loaded through a non-interruption hot replacement technology; A control decision module is used to accumulate the multi-step power prediction values of all production equipment to be operated to obtain a predicted total power curve, and when the predicted total power exceeds a preset power demand threshold, a differentiated regulation strategy is executed according to the control attributes of each production equipment; The differentiated regulation strategy executed by the control decision module comprises: For production equipment with a control attribute of "allowing automatic control", a power-off or delayed start instruction is generated and sent to the corresponding intelligent socket for execution. For the production equipment with control attribute of "key equipment" or "prohibit automatic intervention", a structured early warning information is generated and pushed to the terminal of the preset responsible person.

2. The multi-device power consumption monitoring, prediction and regulation system of claim 1, wherein, The processing process of the work order energy consumption accounting module includes the following steps: Step A1: The work order energy consumption accounting module acquires the current work order information being executed in real time, and combines the real-time state data of the production equipment to screen out a set of target production equipment in the "running" state during the work order operation period; Step A2: For each production equipment in the set of target production equipment, the active power time series data segment of the production equipment during the work order operation period is extracted, and the actual energy consumption value of the production equipment in the work order is calculated: ； In the above formula, represents the first actual energy consumption value of the production equipment within the work order running period; represents the first active power value of the equipment at the first sampling time; represents the sampling time interval; represents the total number of sampling points of the production equipment within the work order running period; Step A3: Accumulate the actual energy consumption values of all the target production devices in the target production device set to obtain the total energy consumption of the work order: the actual energy consumption values of all the target production devices in the target production device set to obtain the total energy consumption of the work order: ； In the above formula, represents the total energy consumption of the work order; Step A4: According to the start and end time of the work order running period, match the time-of-use electricity price strategy, and divide the work order running period into electricity price intervals according to the time dimension, and respectively count the cumulative energy consumption of the target production equipment set in each electricity price interval, and the cumulative energy consumption of the first electricity price interval is recorded as ; Step A5: According to the start and end time of each electricity price interval, the corresponding unit electricity price is obtained, and the unit electricity price of the first electricity price interval is recorded as . ​ Step A6: Accumulate energy consumption of each period Multiply the unit price of each period The total electricity cost of the work order is obtained by accumulating the multiplication results. ； In the above formula, represents the total electricity cost of the work order; Step A7: Read the qualified yield corresponding to the work order from the MES system, divide the total energy consumption of the work order by the qualified yield to obtain the unit product energy consumption, divide the total electricity cost of the work order by the qualified yield to obtain the unit product electricity cost. < 3. The multi-device power consumption monitoring, prediction and regulation system of claim 1, wherein: It also includes a user interaction layer for dynamically presenting the processing results of the edge computing server in a graphical interface.

4. A method for multi-device power consumption monitoring, prediction and regulation, the method comprising: The multi-device power consumption monitoring, prediction and regulation system according to claim 1; The method specifically includes: Step S1: Through the intelligent socket deployed at the front end of the power supply circuit of each production equipment, the voltage, current and power data of the production equipment are collected in real time and uploaded to the edge computing server; Step S2: The work order energy consumption accounting module calculates the unit product energy consumption and unit product electricity cost according to the information of the intelligent socket and the MES system; Step S3: The power prediction module predicts the multi-step power prediction value of each production equipment based on the exclusive model weight; Step S4: The control decision module obtains the predicted total power based on the multi-step power prediction value, compares the predicted total power with the power demand threshold, obtains the control strategy and executes it.

5. The method for multi-device power consumption monitoring, prediction and regulation of claim 4, wherein, It also includes: Real-time analysis of the running state data of each production equipment to identify production equipment that has been in standby state for a long time but has no production task associated with it; If the standby time of a production equipment exceeds the preset threshold, differential treatment is performed according to its control attribute: for non-key equipment with control attribute of "allow automatic control", automatically issue a power-off instruction; for production equipment with control attribute of "key equipment" or "prohibit automatic intervention", generate standby timeout early warning information and push it to the terminal of the relevant responsible person.

Images