An intelligent factory energy management method and system based on big data analysis

Abstract

The application discloses an intelligent factory energy management method and system based on big data analysis, relates to the technical field of intelligent factory energy management, and synchronously collects multi-dimensional energy consumption data through multiple types of intelligent collection terminals deployed in the whole area of the factory, constructs a factory energy big data warehouse after abnormal cleaning and missing interpolation of the edge gateway, extracts a low-dimensional potential vector of an equipment energy consumption mode and a unit product energy consumption dynamic feature, inputs a double-path parallel time series modeling multi-step prediction network to generate an hourly energy consumption demand prediction sequence, combines real-time energy prices to input a multi-agent reinforcement learning scheduling optimization engine to generate an optimal energy scheduling strategy, feeds back execution results to form a closed-loop iterative optimization, and regularly generates an energy management performance report. The scheme comprehensively covers energy consumption data of all links of the factory, improves the accuracy and scene adaptability of energy consumption prediction, provides support for energy management decision-making, and considers energy consumption optimization, carbon emission control and production order stability.

Description

Technical Field

[0001] This invention relates to the field of smart factory energy management technology, and in particular to a smart factory energy management method and system based on big data analysis. Background Technology

[0002] Current manufacturing plants have a wide variety of energy-consuming equipment and complex production line operating conditions, involving energy management across multiple stages including production and auxiliary utilities. Traditional energy management often relies on manual recording of energy consumption data and post-event accounting and statistics, which can only trace past energy consumption but cannot predict future energy demand or make timely adjustments to energy consumption fluctuations. Some factories use simple threshold alarm-based management methods, which only issue alerts when energy consumption exceeds a fixed threshold. They cannot adjust management strategies based on dynamic factors such as production scheduling and environmental changes, which can easily lead to excessively high peak loads and unreasonable allocation of energy resources. They also cannot meet the relevant requirements of carbon emission control and cannot integrate carbon emission control requirements into daily energy scheduling.

[0003] Some existing energy management methods use a single time-series model for energy consumption prediction, analyzing only raw energy consumption time-series data. This fails to distinguish between different types of energy consumption components, such as normal equipment operation, start-up and shutdown fluctuations, and abnormal disturbances. Furthermore, it struggles to fully consider the impact of factors like production output and environmental temperature and humidity on energy consumption, resulting in insufficient matching between predicted results and actual energy demand. Other scheduling optimization schemes employ traditional mathematical programming methods, requiring the pre-establishment of precise physical models of equipment energy consumption. However, adjustments to factory production lines, equipment aging, and changes in operating conditions can all reduce the model's adaptability, making it difficult to cope with dynamically changing production scenarios. Some scheduling schemes using reinforcement learning often employ a single-agent architecture, failing to consider the collaborative needs of different production lines, energy storage systems, and adjustable equipment. This can easily lead to localized optimization without improving overall energy consumption.

[0004] In addition, most existing energy management systems have not formed a complete closed-loop optimization mechanism. The actual effect of the scheduling strategy cannot be automatically fed back to the optimization model for iteration. As the factory's operating status changes, the adaptability of the scheduling strategy will gradually decrease. At the same time, there is a lack of standardized performance statistics mechanism, making it difficult for managers to fully and timely grasp the overall information related to energy management, such as the energy consumption trend and carbon emission of each production line. This makes it impossible to provide sufficient support for subsequent management decisions, and the implementation effect of many energy management measures cannot be evaluated and adjusted in a timely manner. Summary of the Invention

[0005] This invention proposes a smart factory energy management method and system based on big data analysis to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a smart factory energy management method based on big data analysis, comprising the following steps: S1. Energy consumption data such as power consumption of production line equipment, temperature and humidity of workshop, compressed air pressure and three-phase electrical parameters of transformer are collected synchronously through intelligent acquisition terminals. After cleaning and interpolation by the edge gateway, an energy big data warehouse is built according to production line and workshop. S2. Perform variational mode decomposition and multi-layer wavelet packet transform on the time series data of equipment power to obtain intrinsic mode components and sub-band energy, input them into a pre-trained deep convolutional autoencoder, and extract the latent vector of equipment energy consumption mode; use gated recurrent units to extract the output-energy consumption coupling features from the output and energy consumption data to obtain the dynamic feature vector of energy consumption per unit product. S3. Concatenate the latent vector with the dynamic feature vector, input the pre-trained multi-step prediction network, and use a temporal convolutional gating and bidirectional long short-term memory parallel structure to extract short-term fluctuation and long-term trend features, and generate a future hourly energy consumption demand prediction sequence for each production line. S4. Input the demand forecast sequence and real-time energy price into the multi-agent reinforcement learning energy scheduling optimization engine. Use adjustable equipment power, energy storage charging and discharging power, production task scheduling deviation and demand response as actions, and use the weighted sum of energy cost and carbon emissions as the optimization objective. Use multi-agent proximal strategy optimization for offline pre-training and online fine-tuning, and output equipment parameter adjustment, energy storage plan and scheduling optimization strategy. S5. After execution, the results will be fed back to the experience playback buffer, and energy management performance reports will be generated regularly, including the energy consumption trend of each production line unit, the time-of-use energy cost composition, and the carbon emission compliance rate.

[0007] Furthermore, it also includes: in step S1, the environmental temperature and humidity data of each workshop and the real-time power data of each production line equipment are verified by Granger causality test and transfer entropy analysis to complete the dual causal relationship verification, identify the significant influence channels and lag orders of environmental temperature and humidity on equipment energy consumption, and incorporate the temperature and humidity features and lag orders corresponding to the significant influence channels as exogenous inputs into the input feature set of the factory energy consumption multi-step prediction network described in step S3. In step S4, the multi-agent near-end policy optimization applies a trust domain constraint to the KL divergence of the new and old policies of each agent as a unified upper limit of the global policy update magnitude.

[0008] Furthermore, it also includes: when constructing the factory energy scheduling optimization engine in step S4, a multi-agent collaborative communication module based on attention mechanism is introduced. Each agent uses its own observation state as the query vector and the observation states of other agents as the key and value vectors in the decision step. The collaborative attention weight is calculated by scaling dot product attention. The weighted aggregated information of other agents is concatenated with its own observation state as the input of the policy network.

[0009] Furthermore, the joint time-frequency feature extraction described in step S2 is performed as follows: variational mode decomposition is first performed on the power time series data of each production line equipment to obtain a preset number of intrinsic mode components with different center frequencies. Then, wavelet packet transform with a preset number of decomposition layers is performed on each intrinsic mode component to obtain the normalized energy distribution vector of each component in each sub-frequency band. The sub-frequency band energy distribution vectors of all intrinsic mode components are concatenated in the order of component index and then input into a deep convolutional autoencoder.

[0010] Furthermore, in step S3, the temporal convolutional gating branch of the factory energy consumption multi-step prediction network is composed of stacked pre-specified number dilated causal convolutional layers, with the dilation factor of each layer increasing exponentially by a pre-specified base. The bidirectional long short-term memory branch is composed of forward and backward long short-term memory layers connected in parallel. The two hidden states are concatenated by time steps and mapped to the predicted values ​​through a fully connected layer. The two outputs are processed by a learnable gating fusion unit to generate fusion weights for each time step to complete adaptive weighted fusion, and then processed by an output fully connected layer to generate an hourly energy consumption prediction sequence.

[0011] Furthermore, the weighted combination optimization objective of the factory's comprehensive energy cost and carbon emissions mentioned in step S4 is defined by the following formula: ; In the formula, Indicates that the factory contains The comprehensive optimization objective value within the scheduling cycle of each decision time step. For decision time step index, As a weighted factor for energy costs, For carbon emission weights, Assign task offset penalty weights, Indicates the first Factory energy costs at each time step Indicates the first Factory carbon emissions at each time step Indicates the first A vector of production task scheduling offsets for each time step. express Norm.

[0012] Furthermore, the factory energy management performance report mentioned in step S5 is automatically generated at the end of each preset analysis cycle. The report includes the current cycle unit product energy consumption trend of each production line and the percentage deviation compared with the same period in history and industry benchmarks, the composition of the factory's total energy cost by time period and the correlation analysis of the cost ratio of each time period with the peak and valley electricity price distribution, the ratio of the factory's total carbon emissions and carbon emission compliance rate in the current cycle to the government-approved quota, and the actual adoption rate of each optimization suggestion in the scheduling strategy of the previous cycle and the energy saving quantitative assessment after adoption.

[0013] Furthermore, a smart factory energy management system based on big data analytics includes the following modules: The multi-source energy data acquisition and aggregation module is equipped with a cluster of intelligent acquisition terminals distributed in various workshops of the factory and an anomaly cleaning and missing data interpolation gateway deployed at the edge. It is used to collect multi-dimensional energy consumption data across the entire factory and build a factory energy big data warehouse according to production line and workshop indexes. The joint energy consumption feature extraction module is equipped with a variational mode decomposition unit, a wavelet packet transform unit, a deep convolutional autoencoder unit, and a production value-energy consumption coupled feature extraction unit, which is used to extract low-dimensional potential vectors of energy consumption patterns and dynamic feature vectors of energy consumption per unit product from equipment power data. The energy consumption multi-step time series prediction module is equipped with a parallel temporal convolutional gated network branch, a bidirectional long short-term memory network branch, and a learnable gated fusion unit, which is used to generate an hourly energy consumption demand prediction sequence for each production line within a future preset time period. The multi-agent energy scheduling optimization module is equipped with an attention-based collaborative communication unit, a multi-agent proximal policy optimization trainer, and an experience replay buffer, which are used to output the optimal energy scheduling strategy covering equipment parameter adjustment, energy storage planning, and scheduling optimization. The scheduling execution closed-loop and performance reporting module is equipped with an actual execution result feedback unit and an automatic performance report generation unit. It is used to feed the execution results back to the optimization engine and regularly generate energy management performance reports containing energy consumption trends, cost structure and carbon emission compliance rate.

[0014] Furthermore, the multi-agent energy scheduling optimization module models each adjustable device and each energy storage system as an independent reinforcement learning agent. The observation space of each agent includes the current power state of its device, the upper and lower limits of the adjustable range, and the time label of the local electricity price. The policy network adopts a three-layer fully connected hidden layer plus a Gaussian policy output layer structure. The mean of the output action is mapped to the adjustable percentage range of the rated power by tanh, and the exploration amplitude is controlled by the logarithm of the action standard deviation. The experience replay buffer of each agent is maintained independently. Global training is carried out by the central coordinator by extracting an equal number of samples from each buffer at a preset period to carry out small-batch gradient updates.

[0015] Furthermore, the learnable gated fusion unit of the energy consumption multi-step time series prediction module concatenates the output features of the temporal convolutional gated branch and the output features of the bidirectional long short-term memory branch along the time step dimension, and obtains the fusion weights of the two branches at each time step through a single-layer fully connected layer and sigmoid activation mapping. The weighted fusion feature vectors of each time step are input to the fully connected layer to generate an hourly energy consumption prediction sequence.

[0016] Compared with existing technologies, the beneficial effects of this invention are: This solution achieves unified and synchronized collection and standardized preprocessing of multi-dimensional energy data across the entire factory, covering energy-related data from all stages, including production workshops and auxiliary facilities. This provides comprehensive foundational data support for subsequent energy consumption analysis, prediction, and scheduling. By employing a combination of joint time-frequency feature extraction and deep convolutional autoencoders to extract equipment energy consumption pattern features, it can separate energy consumption components with different characteristics, more accurately reflecting the actual operating energy consumption patterns of equipment. Furthermore, by combining output-energy consumption coupling feature extraction, it can correlate production status with energy consumption data, avoiding the problem of analyzing only energy consumption data while ignoring actual production needs.

[0017] The energy consumption prediction module of this solution adopts a dual-path parallel time-series modeling structure, extracting short-term fluctuations and long-term trend features of energy consumption. It can incorporate causally validated environmental exogenous features, adapting to energy consumption changes under different operating conditions and seasons, and improving the matching degree between prediction results and actual energy demand. The scheduling optimization part adopts a multi-agent reinforcement learning architecture, which does not require the pre-establishment of accurate equipment physical models. It can complete training and iteration based on actual operating data. The added multi-agent collaborative communication module enables each agent to perceive the operating status and energy consumption plan of other equipment and energy storage when making decisions, avoiding the overall load imbalance problem caused by local adjustment. The trust domain constraint applied to policy updates makes the policy update process more stable and avoids excessive policy fluctuations. The optimization objective simultaneously covers energy consumption, carbon emission-related indicators and production scheduling deviation penalties, which can reduce the impact of scheduling strategies on normal production order while optimizing energy consumption and carbon emission-related indicators.

[0018] This solution also includes a closed-loop optimization mechanism. The actual execution results of the scheduling strategy can be fed back to the experience playback buffer of the scheduling optimization engine, enabling continuous iterative optimization of the scheduling model. This allows the solution to adapt to changes in operating status, such as factory production line adjustments and equipment aging. The regularly generated energy management performance reports can cover multi-dimensional information such as energy consumption trends, energy consumption structure, and carbon emissions, providing managers with comprehensive references to support the formulation of subsequent energy management-related decisions and facilitating the evaluation of the effectiveness of implemented energy management measures. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the intelligent factory energy management method based on big data analysis proposed in this invention; Figure 2 This is a flowchart of the multi-source energy data acquisition and aggregation process of the present invention; Figure 3 This is a flowchart of the joint extraction and coupling of energy consumption features in this invention; Figure 4 This is a flowchart of the multi-step energy consumption timing prediction process with integrated gating in this invention; Figure 5This is a flowchart of the multi-agent collaborative energy scheduling and closed-loop optimization process of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0023] Reference Figures 1 to 5 A smart factory energy management method based on big data analytics includes the following steps: S1. By deploying various types of intelligent acquisition terminals in production workshops and public auxiliary facilities in the factory, real-time power of each production line equipment, temperature and humidity of each workshop, pressure of compressed air pipeline network, and three-phase electrical parameters of the low-voltage side of the transformer are collected synchronously with a unified time reference. After anomaly cleaning and missing interpolation are performed by the edge gateway, the factory energy big data warehouse is built according to the production line and workshop index. S2. Perform joint time-frequency feature extraction based on variational mode decomposition and multi-layer wavelet packet transform on the real-time power time-series data of each production line equipment. Input the energy values ​​of each intrinsic mode component and sub-band after decomposition into a pre-trained deep convolutional autoencoder to extract the low-dimensional latent vector of equipment energy consumption mode. Perform output-energy consumption coupling feature extraction based on gated cyclic unit on the product output and energy consumption data of each production line to obtain the dynamic feature vector of unit product energy consumption. S3. The low-dimensional potential vector of equipment energy consumption mode and the dynamic feature vector of unit product energy consumption are concatenated and input into the pre-trained factory energy consumption multi-step prediction network. The network extracts the short-term fluctuation features and long-term trend features of energy consumption through the dual-path parallel temporal modeling structure of temporal convolutional gating and bidirectional long short-term memory, and then generates the hourly energy consumption demand prediction sequence of each production line in the future preset period. S4. Input the hourly energy demand forecast sequence and the factory's real-time energy price signal into the factory energy scheduling optimization engine based on multi-agent reinforcement learning. The action space is the power setpoint of each adjustable equipment, the energy storage charging and discharging power, the scheduling offset of transferable production tasks, and the demand response participation. The optimization objective is the weighted combination of the factory's comprehensive energy cost and carbon emissions. The multi-agent near-end strategy optimization algorithm is pre-trained offline on historical data and fine-tuned online to output the optimal energy scheduling strategy covering equipment parameter adjustment, energy storage planning, and scheduling optimization. S5. After executing the optimal energy dispatch strategy, the actual execution results are fed back to the experience playback buffer of the dispatch optimization engine to form a closed-loop iterative optimization, and a factory energy management performance report containing the energy consumption trend of each production line unit product, the time-of-use energy cost composition and the carbon emission compliance rate is generated according to the specified period.

[0024] The present invention further includes: in step S1, verifying the dual causal relationship between the environmental temperature and humidity data of each workshop and the real-time power data of each production line equipment based on Granger causality test and transfer entropy analysis, identifying the significant influence channels of environmental temperature and humidity on equipment energy consumption and their lag order, and incorporating the temperature and humidity features and lag order corresponding to the significant influence channels as exogenous inputs into the input feature set of the factory energy consumption multi-step prediction network described in step S3 to improve the adaptability to seasonal environmental changes and changes in equipment heat dissipation efficiency; in step S4, multi-agent proximal policy optimization applies a trust domain constraint to the KL divergence of the new and old policies of each agent as a unified upper limit for the global policy update magnitude.

[0025] This invention also includes: introducing a multi-agent collaborative communication module based on an attention mechanism during the construction of the factory energy scheduling optimization engine in step S4. Each agent uses its own observation state as a query vector and the observation states of other agents as key and value vectors to calculate the collaborative attention weight through scaling dot product attention in the decision step. The weighted aggregated information of other agents is concatenated with its own observation state as the input of the policy network, so that each agent can perceive the energy consumption plan and energy storage status of other production line equipment when making power adjustment decisions, thereby avoiding load imbalance caused by local adjustment.

[0026] In this invention, the joint time-frequency feature extraction in step S2 is performed as follows: variational mode decomposition is first performed on the power time-series data of each production line equipment to obtain a preset number of intrinsic mode components with different center frequencies, so that the different frequency components correspond to the three components of normal operation power, periodic start-stop fluctuation and random abnormal fluctuation respectively. Then, wavelet packet transform with a preset number of decomposition layers is performed on each intrinsic mode component to obtain the normalized energy distribution vector of each component in each sub-frequency band. The sub-frequency band energy distribution vectors of all intrinsic mode components are concatenated in the order of component index and then input into the deep convolutional autoencoder.

[0027] In this invention, the temporal convolutional gating branch of the factory energy consumption multi-step prediction network in step S3 is composed of a preset number of dilated causal convolutional layers stacked together, and the dilation factor of each layer increases in a preset power of a preset base to expand the temporal receptive field. The bidirectional long short-term memory branch is composed of forward and backward long short-term memory layers connected in parallel, and the two hidden states are concatenated by time steps and mapped to the predicted value through a fully connected layer. The two outputs are processed by a learnable gating fusion unit to generate fusion weights for each time step, and then adaptively weighted and fused before being processed by an output fully connected layer to generate an hourly energy consumption prediction sequence.

[0028] In this invention, the weighted combination optimization objective of the factory's comprehensive energy cost and carbon emissions in step S4 is defined by the following formula: ; In the formula, Indicates that the factory contains The comprehensive optimization target value within the scheduling cycle of each decision time step, with the dimension being a weighted sum of monetary units and carbon emission units; The decision time step index is an integer ranging from 1 to... ; The energy cost is weighted and dimensionless. This is a carbon emission weight with the dimension of carbon emission monetization coefficient. The task offset penalty weight is expressed as the offset cost per hour in monetary units. Indicates the first The energy cost of the factory at each time step, expressed in monetary units; Indicates the first The carbon emissions of the factory at each time step, measured in tons of carbon emissions; Indicates the first A time-step production task scheduling offset vector with the dimension of production man-hours; express Norm.

[0029] In this invention, the factory energy management performance report mentioned in step S5 is automatically generated at the end of each preset analysis cycle. The report includes the current cycle unit product energy consumption trend of each production line and the percentage deviation compared with the same period in history and industry benchmarks, the composition of the factory's total energy cost by time period and the correlation analysis of the cost ratio of each time period with the peak and valley electricity price distribution, the ratio of the factory's total carbon emissions and carbon emission compliance rate in the current cycle to the government-approved quota, and the actual adoption rate of each optimization suggestion in the optimal scheduling strategy of the previous cycle and the energy saving quantitative assessment after adoption.

[0030] This invention discloses an intelligent factory energy management system based on big data analytics, comprising the following modules: The multi-source energy data acquisition and aggregation module is equipped with a cluster of intelligent acquisition terminals distributed in various workshops of the factory and an anomaly cleaning and missing data interpolation gateway deployed at the edge. It is used to collect multi-dimensional energy consumption data across the entire factory and build a factory energy big data warehouse according to production line and workshop indexes. The joint energy consumption feature extraction module is equipped with a variational mode decomposition unit, a wavelet packet transform unit, a deep convolutional autoencoder unit, and a production value-energy consumption coupled feature extraction unit, which is used to extract low-dimensional potential vectors of energy consumption patterns and dynamic feature vectors of energy consumption per unit product from equipment power data. The energy consumption multi-step time series prediction module is equipped with a parallel temporal convolutional gated network branch, a bidirectional long short-term memory network branch, and a learnable gated fusion unit, which is used to generate an hourly energy consumption demand prediction sequence for each production line within a future preset time period. The multi-agent energy scheduling optimization module is equipped with an attention-based collaborative communication unit, a multi-agent proximal policy optimization trainer, and an experience replay buffer, which are used to output the optimal energy scheduling strategy covering equipment parameter adjustment, energy storage planning, and scheduling optimization. The scheduling execution closed-loop and performance reporting module is equipped with an actual execution result feedback unit and an automatic performance report generation unit. It is used to feed the execution results back to the optimization engine to form a closed loop and periodically generate energy management performance reports containing energy consumption trends, cost structure and carbon emission compliance rate.

[0031] In this invention, the multi-agent energy scheduling optimization module models each adjustable device and each energy storage system as an independent reinforcement learning agent. The observation space of each agent includes the current power state of the device, the upper and lower limits of the adjustable range, and the time label of the local electricity price. The policy network adopts a structure of three fully connected hidden layers plus a Gaussian policy output layer. The mean of the output action is mapped to the adjustable percentage range of the rated power by tanh and the exploration amplitude is controlled by an independent output head of the logarithm of the action standard deviation. The experience replay buffer of each agent is maintained independently and the global training is performed by the central coordinator by extracting an equal number of samples from each buffer at a preset period for small-batch gradient updates.

[0032] In this invention, the learnable gated fusion unit of the energy consumption multi-step time-series prediction module concatenates the output features of the temporal convolution gated branch and the output features of the bidirectional long short-term memory branch in the time-step dimension, and then maps them to the fusion weights of the two branches in each time step through a single-layer fully connected layer and sigmoid activation. After weighted fusion, the feature vectors of each time step are input to the fully connected layer to generate an hourly energy consumption prediction sequence. The parameters of the learnable gated fusion unit are optimized through end-to-end backpropagation training.

[0033] The specific embodiments of the present invention are further illustrated below: The first embodiment is applied to a discrete manufacturing smart factory for metal structural parts processing. The factory is set up with four production workshops: blanking, welding, machining and surface treatment, and is equipped with public auxiliary facilities such as air compressor station, substation and refrigeration station.

[0034] The multi-source energy data acquisition and aggregation module deploys corresponding acquisition terminals within the factory area. Electrical parameter acquisition terminals are installed at the inlet terminals of each production line equipment; temperature and humidity sensing terminals are installed at representative points in each workshop's production area; pressure sensing terminals are installed at the main nodes and branch ends of the compressed air pipeline network; and three-phase electrical parameter acquisition terminals are installed at the low-voltage side outlet terminals of each transformer. All acquisition terminals are uniformly connected to the factory time synchronization server, synchronously acquiring multi-dimensional energy-related data across the entire area according to a unified time reference and a set sampling frequency. The acquired data includes real-time electrical power of each production line equipment, ambient temperature and humidity in each workshop, compressed air pipeline network pressure, and three-phase electrical parameters on the low-voltage side of the transformers. The raw data is transmitted to a gateway deployed at the edge. The gateway first identifies and marks outliers in the data using the 3σ criterion, then combines the moving average of adjacent acquisition points within the same time window with the concurrent operating data of similar equipment in the same workshop to imputate missing values. The pre-processed data is indexed in a two-layer system according to production line number and workshop number and stored in the factory's energy big data warehouse. After the data is entered into the warehouse, the system conducts dual causal relationship verification on the environmental temperature and humidity data of each workshop and the real-time power data of the corresponding production line equipment in the workshop. First, the Granger causality test is used to preliminarily screen out the channels with statistical causality between the temperature and humidity sequence and the power sequence. Then, the information transmission amount and corresponding lag order of each channel are determined by the transfer entropy calculation. The influence channels with information transmission amount higher than the set threshold are screened out, and the corresponding temperature and humidity features and lag order are organized into exogenous input features for later use.

[0035] The energy consumption feature extraction module extracts features according to a set process. For the power time-series data of each production line device, variational mode decomposition is first performed. The number of intrinsic mode components is preset according to the device type. The different center frequency components obtained by decomposition correspond to the stable power components of normal device operation, the periodic fluctuation components caused by periodic device start-stop and workstation switching, and the random abnormal components caused by device no-load disturbance or power supply fluctuation. After completing the variational mode decomposition, multi-level wavelet packet transform is performed on each intrinsic mode component. The number of decomposition levels is set according to the data sampling frequency. The normalized energy value of each component in different sub-frequency bands is calculated. The sub-frequency band energy distribution vectors corresponding to all intrinsic mode components of the same device are concatenated in order of component frequency from low to high to obtain the high-dimensional time-frequency features of device energy consumption. The deep convolutional autoencoder consists of an encoder and a decoder. The encoder comprises four stacked convolutional layers, each followed by a batch normalization layer and a ReLU activation layer. The kernel size decreases layer by layer, while the number of channels increases layer by layer, mapping the high-dimensional time-frequency features of the input to a fixed-dimensional low-dimensional latent vector. The decoder consists of four stacked transposed convolutional layers, symmetrical to the encoder, used to reconstruct the low-dimensional latent vector back into the original high-dimensional time-frequency features. During model training, the time-frequency features of equipment under normal operating conditions stored in an energy big data warehouse are used as the training set. The training and validation sets are divided in an 8:2 ratio. The training objective is to minimize the reconstruction error. The Adam optimizer is used for parameter iteration. Once the reconstruction error on the validation set stabilizes within a preset range, the model parameters are fixed. The encoder structure is retained for real-time feature extraction, outputting a low-dimensional latent vector of the device's energy consumption mode. The output-energy consumption coupling feature extraction adopts a gated recurrent unit network with a two-layer hidden layer structure. The input is the hourly qualified output data of each production line aligned by time step and the total energy consumption data of the corresponding time period. The network controls the transmission of time sequence information through update gate and reset gate. After training, it outputs the dynamic feature vector of unit product energy consumption of the corresponding time period.

[0036] The multi-step time-series energy consumption prediction module receives the concatenated low-dimensional latent vector of equipment energy consumption patterns and the dynamic feature vector of unit product energy consumption. Simultaneously, it incorporates pre-processed exogenous features of environmental temperature and humidity. All features are aligned to the same time step according to their corresponding lag orders and input into the factory's multi-step energy consumption prediction network. The network employs a dual-path parallel temporal modeling structure. The temporal convolutional gating branch consists of multiple stacked dilated causal convolutional layers, with the dilation factor increasing progressively in powers of 2. This progressively expanding temporal receptive field extracts rapid energy consumption fluctuations within short time windows. Each dilated causal convolutional layer is followed by a gating linear unit to control feature transmission. The bidirectional long short-term memory branch consists of a forward long short-term memory layer and a backward long short-term memory layer connected in parallel. The forward layer transmits temporal information in forward chronological order, while the backward layer transmits it in reverse chronological order. The hidden states output by both paths at each time step are concatenated to extract energy consumption trend features within a long time window. The learnable gated fusion unit concatenates the features output from the two branches along the time step dimension. It calculates the fusion weights of the two branches at each time step using a single fully connected layer and a sigmoid activation function. The two features are then adaptively weighted and fused according to these weights. The fused features are input to and output to the fully connected layer, generating hourly energy consumption demand prediction sequences for each production line over the next 24 hours. The prediction network is trained using three years of historical energy consumption, output, and environmental feature data from the factory. The training objective is to minimize the average absolute error between the predicted and actual energy consumption sequences. End-to-end backpropagation training is performed using the AdamW optimizer. During training, a cosine annealing strategy is used to adjust parameters and update the step size. Once the prediction error on the validation set stabilizes, the model parameters are fixed for real-time prediction.

[0037] The multi-agent energy scheduling optimization module models each adjustable-power production device and each energy storage system as an independent reinforcement learning agent. Each agent's observation space includes the current operating power of its device, the upper and lower limits of its adjustable power range, the energy price tag for the current time period, and the load status tag for its region. The policy network for each agent adopts a three-layer fully connected hidden layer structure. The hidden layer activation function is ReLU, and the output layer is a Gaussian policy output layer with two independent output heads. One output head outputs the action mean, which, after tanh activation, is mapped to the adjustable percentage range of the rated power, corresponding to the power adjustment amplitude of the device or the direction and magnitude of the charging and discharging power of the energy storage. The other output head outputs the logarithm of the action standard deviation, used to control the exploration amplitude during training and online operation. Before making a decision, each agent interacts with other agents through a collaborative communication module based on an attention mechanism. Each agent uses its own observed state vector as a query vector and the observed state vectors of all other agents as key and value vectors. A scaled dot product attention algorithm is used to calculate the collaborative attention weights for the states of other agents. The state information of other agents is then weighted and aggregated according to these weights. The aggregated information is then concatenated with the agent's own observed state vector and input into its own policy network. Policy training employs a multi-agent proximal policy optimization algorithm, divided into two stages: offline pre-training and online fine-tuning. In the offline pre-training stage, a simulation environment is constructed using historical operating data stored in an energy big data warehouse to train the policy and value networks of each agent. During training, a uniform trust domain upper limit is set for the KL divergence between the old and new policies of each agent to limit the magnitude of each policy update. The optimization process comprehensively considers the factory's energy costs within the scheduling cycle, the carbon emissions from corresponding energy consumption, and the impact of scheduling offsets of transferable production tasks. These three components are combined according to preset weights as the optimization direction. The agent's action space includes the power setpoint adjustment values ​​of each adjustable device, the charging and discharging power values ​​of the energy storage system, the scheduling offset of transferable production tasks, and the participation of demand response projects. After training, the system enters the online operation phase. During online operation, the model parameters are fine-tuned using real-time collected data, and the optimal energy scheduling strategy for the current scheduling cycle is finally output, covering equipment operating parameter adjustment schemes, energy storage time-sharing charging and discharging plans, production task scheduling optimization schemes, and demand response participation schemes.

[0038] After the scheduling strategy is issued and executed, the scheduling execution closed-loop and performance reporting module collects actual operational data from each stage, including actual power adjustment values, actual charging and discharging amounts, actual scheduling adjustments, actual energy costs, and actual carbon emission data. This data is then organized according to time steps and stored in experience playback buffers independently maintained by each agent. The central coordinator extracts an equal number of samples from each buffer at fixed intervals for small-batch gradient updates, forming a closed-loop optimization mechanism. After each preset analysis cycle, the system automatically generates an energy management performance report. The report includes the unit product energy consumption trend of each production line within the statistical period, as well as the comparison deviation with historical data and industry benchmark reference values. It also analyzes the composition of the factory's total energy costs by time period, the correlation between the cost proportion of each time period and the time-of-use electricity price distribution, the factory's total carbon emissions within the statistical period, and the ratio of carbon emission compliance rate to the approved quota. Simultaneously, it statistically analyzes the actual adoption of scheduling optimization suggestions output in the previous cycle and quantitatively evaluates the energy-saving effects of the adopted suggestions.

[0039] This embodiment is applied to discrete manufacturing plants that process metal structural parts. It can cover the collection and standardized processing of energy consumption data for all production links and public auxiliary facilities in the entire plant area. Through multi-dimensional feature extraction, it fully reflects the coupling relationship between the actual operating energy consumption mode of the equipment and the production energy consumption. The dual-path parallel prediction structure can adapt to the characteristics of frequent equipment start-ups and shutdowns and large energy consumption fluctuations in discrete manufacturing scenarios. Multi-agent collaborative scheduling can balance the energy demand of each workshop and each piece of equipment, avoiding load fluctuations caused by local power adjustments. The closed-loop iterative mechanism can adapt to the drift of energy characteristics caused by order switching and changes in equipment operating conditions in discrete manufacturing scenarios. The automatically generated performance report can help managers fully grasp the energy operation status of the plant area and support the implementation of energy management decisions.

[0040] The second embodiment is applied to a smart factory for printed circuit board assembly process manufacturing. The factory is set up with four continuous production workshops: SMT placement, through-hole soldering, assembly testing, and packaging and warehousing. It is equipped with supporting public auxiliary facilities such as pure water station, air compressor station, central air conditioning station, and substation. The production process is highly continuous and the overall energy load is stable, but it is significantly affected by the ambient temperature and humidity and the adjustment of the production cycle.

[0041] The multi-source energy data acquisition and aggregation module deploys acquisition terminals at corresponding locations within the factory area. Electrical parameter acquisition terminals are installed at the power supply nodes of each continuous production line equipment; temperature and humidity sensing terminals are installed in environmentally sensitive production workstations in each workshop; pressure sensing terminals are installed at key nodes in the compressed air and pure water supply networks; and three-phase electrical parameter acquisition terminals are installed at the low-voltage side output terminals of each transformer. All acquisition terminals are connected to the factory's unified time synchronization system, synchronously collecting multi-dimensional energy-related data across the entire area at a fixed sampling frequency. The collected data includes real-time electrical power of each production line equipment, ambient temperature and humidity in each workshop, compressed air network pressure, and three-phase electrical parameters on the low-voltage side of the transformers. The raw data is transmitted to the edge-side gateway for preprocessing. Anomaly cleaning uses a sliding window threshold combined with equipment operating status marking to identify outliers. Abnormal power fluctuations collected when equipment is stopped and jump values ​​caused by communication interruptions of sensing terminals are marked as invalid. Missing data is interpolated using synchronous data from adjacent workstations on the same production line combined with time series interpolation. After preprocessing, the data is indexed by workshop number and production line number and stored in the factory's energy big data warehouse. After the data is stored in the warehouse, considering the significant impact of ambient temperature and humidity on air conditioning and equipment heat dissipation energy consumption in the production process, the system conducts Granger causality tests and transfer entropy analysis on the temperature and humidity sequences of each workshop and the corresponding power sequences of production line equipment. First, it screens out the correlation channels with statistical causality, then calculates the information transmission strength and corresponding lag time of each channel, and screens out the influential channels whose information transmission strength reaches the judgment threshold. The corresponding temperature and humidity features and lag order are organized into exogenous input features to provide supplementary input for subsequent energy consumption prediction.

[0042] The energy consumption feature extraction module extracts features based on the energy consumption characteristics of the manufacturing process. It performs variational mode decomposition on the power time-series data of each production line device, setting a fixed number of intrinsic mode components. The different center frequency components obtained from the decomposition correspond to the baseline power component for continuous and stable production, the periodic fluctuation component caused by production cycle adjustments and model changes, and the random fluctuation component caused by grid disturbances and temporary equipment idleness. After completing the variational mode decomposition, a multi-level wavelet packet transform is performed on each intrinsic mode component. The number of decomposition levels is set according to the data sampling frequency, and the normalized energy value of each sub-band is calculated. The energy vectors of all intrinsic mode components corresponding to the sub-bands are concatenated in ascending order of component frequency to obtain high-dimensional time-frequency features. The deep convolutional autoencoder consists of four convolutional layers at the encoder end, each followed by a batch normalization layer and a ReLU activation layer. The number of channels increases progressively with each layer, and the kernel size adapts to the continuous characteristics of the time-series data. The decoder end consists of four symmetrical transposed convolutional layers. During training, the factory's normal operation time-frequency feature data for three consecutive years is used as the training set, divided into training and validation sets in a 7:3 ratio. The Adam optimizer is used for iterative training with the goal of minimizing reconstruction error. Training stops when the reconstruction error of the validation set stabilizes within a preset range. The encoder network is retained to map the high-dimensional time-frequency features of the input to a low-dimensional latent vector of equipment energy consumption patterns. The output-energy consumption coupling feature extraction uses a gated recurrent unit network with a two-layer hidden layer structure. The input is the number of qualified circuit boards produced by each production line hourly, aligned to the time step, and the total energy consumption data of the production line for the corresponding time period. The gating mechanism filters out effective correlation information under long time series. After training, the output is a dynamic feature vector of energy consumption per unit product, reflecting the coupling relationship between changes in production capacity and energy consumption during continuous production.

[0043] The multi-step time-series energy consumption prediction module concatenates the low-dimensional latent vector of equipment energy consumption patterns with the dynamic feature vector of unit product energy consumption in the feature dimension. Simultaneously, it incorporates processed exogenous features of temperature and humidity, aligning all features to the same time step according to their corresponding lag order before inputting them into the factory energy consumption multi-step prediction network. The network employs a dual-path parallel structure. The temporal convolutional gated branch consists of stacked dilated causal convolutional layers, with the dilation factor increasing progressively in powers of 2. This continuously expands the receptive field, capturing rapid fluctuations in energy consumption within short time windows. Each convolutional layer is followed by a gated linear unit to filter out invalid features. The bidirectional long short-term memory branch consists of two parallel long short-term memory layers (forward and backward), extracting long-term trend features of energy consumption in the continuous production process from both forward and reverse time directions. The hidden states from both directions are concatenated at each time step. The two outputs are fed into a learnable gated fusion unit, where the two features are concatenated along the time step dimension. The fusion weights for each time step are then calculated using a single fully connected layer and sigmoid activation. Adaptive weighted fusion is performed according to these weights, and the fused features are fed into the output fully connected layer to generate hourly energy consumption demand prediction sequences for each production line over the next 48 hours. The prediction network is trained using four years of continuous historical operating data from the factory as the dataset. The training objective is to minimize the mean squared error between the predicted and actual energy consumption sequences. An AdamW optimizer is used for end-to-end backpropagation parameter updates. An early stopping mechanism is employed during training to prevent overfitting; training stops when the prediction error on the validation set no longer decreases after several consecutive rounds, fixing the model parameters for online prediction.

[0044] The multi-agent energy scheduling optimization module, designed for the continuous nature of process manufacturing, models the adjustable process equipment of each production line, the adjustable load of each public auxiliary system, and the energy storage system configured in the plant as independent reinforcement learning agents. Each agent's observation space includes the current operating power of its equipment or system, the upper and lower limits of its adjustable power range, the current electricity price period, and the remaining capacity of its power supply path. The policy network for each agent employs a three-layer fully connected hidden layer with ReLU activation. The output layer is a Gaussian policy output layer, outputting the action mean and the logarithm of the action standard deviation. The action mean is mapped to the adjustable percentage range of the corresponding equipment's rated power using tanh activation, while the action standard deviation is used to control the exploration amplitude during training and operation. Agents interact with each other through a collaborative communication module based on an attention mechanism. Each agent uses its own observed state as a query vector and the observed states of other agents as key and value vectors. Collaborative weights are calculated using scaled dot product attention. The weighted aggregation of the state information of other agents is then concatenated with the agent's own observed state as input to the policy network. This allows agents to perceive energy usage plans in other processes during decision-making, preventing overall energy supply instability caused by localized load adjustments during continuous production. Policy training employs a multi-agent near-end policy optimization algorithm. Offline pre-training is first conducted using a simulation environment built based on historical data from an energy big data warehouse. During training, a unified upper limit for the KL divergence trust domain is set between the old and new policies to constrain the magnitude of each parameter update and ensure training stability. The optimization process comprehensively considers the overall energy cost of the factory within the scheduling cycle, the carbon emissions corresponding to energy consumption, and the impact of transferable production task scheduling offsets on the continuous production rhythm. Pre-defined weight combinations are used as optimization directions. The agent's action space includes adjusting the power setpoint of adjustable equipment, setting the charging and discharging power of energy storage, setting the scheduling offset of transferable production tasks, and setting the grid demand response participation. After offline pre-training, the model is fine-tuned online by incorporating real-time operational data, and outputs the optimal energy dispatch strategy covering equipment parameter adjustment, energy storage charging and discharging plans, scheduling optimization, and demand response participation.

[0045] After the scheduling strategy is issued to the control systems of each equipment, the energy storage management system, and the production scheduling system, the scheduling execution closed-loop and performance reporting module collects the actual operating data of each link, including the actual operating power of the equipment, the actual charging and discharging of the energy storage, the actual adjustment of the production tasks, the actual energy cost generated, and the actual carbon emissions. After organizing this data according to the time step, it is stored in the experience playback buffer maintained independently by each agent. The central coordinator extracts an equal number of samples from the experience playback buffer of each agent at a fixed period and performs small-batch gradient updates on the policy network and the value network to achieve continuous closed-loop iterative optimization of the strategy. At each preset analysis cycle node, the system automatically generates an energy management performance report. The report includes the unit product energy consumption change trend of each production line within the statistical cycle, as well as the deviation ratio compared with the historical level and industry benchmark reference value. It also calculates the energy cost composition of the entire factory by time period, analyzes the relationship between the cost ratio of each time period and the distribution of time-of-use electricity prices, calculates the total carbon emissions of the entire factory within the statistical cycle and the ratio of carbon emission compliance rate to the government-approved quota, and calculates the actual adoption rate of scheduling optimization suggestions from the previous cycle, and quantitatively evaluates the energy-saving effect after adopting the suggestions.

[0046] This embodiment is applied to a process manufacturing plant for printed circuit board assembly. It can adapt to the energy consumption characteristics of continuous production scenarios, such as gradual changes in energy consumption, obvious long-term trends, and direct impact from environmental factors. By extracting joint time-frequency features, it can separate the baseline energy consumption, fluctuating energy consumption, and disturbance components in continuous production. The dual-path parallel prediction model can accurately reflect the energy consumption change trend of long-cycle production while capturing short-term energy consumption fluctuations. The multi-agent scheduling with attention communication mechanism can ensure the balance of energy supply and consumption in the continuous production process and avoid the impact of load adjustment on production stability. The closed-loop iteration mechanism can adapt to the changes in energy consumption characteristics caused by production process iteration and slow drift of equipment performance. The regularly generated performance reports can support the routine energy management work in process manufacturing scenarios.

[0047] Reference Figure 1 The intelligent factory energy management method of this invention covers a complete closed loop from bottom-level data acquisition to top-level decision optimization. The system first utilizes various terminal nodes to synchronously collect energy consumption and environmental data across the entire factory and establishes a unified energy big data warehouse. Subsequently, it conducts in-depth joint time-frequency feature analysis on the collected high-dimensional data to deconstruct the potential energy consumption patterns and output coupling characteristics of the equipment. Based on feature extraction, the system uses a dual-path parallel time-series network to accurately predict the future energy demand of each production line. Based on this demand prediction sequence, a multi-agent scheduling engine comprehensively considers energy costs and carbon emission constraints, autonomously seeks optimization, and issues optimal energy scheduling commands. The final execution results continuously flow back to the system to complete reinforcement learning iterations and periodically generate multi-dimensional energy management performance reports.

[0048] Reference Figure 2The multi-source energy data acquisition and aggregation stage is a crucial step in building the foundation for intelligent decision-making. Sensor terminal arrays distributed throughout the factory workshops are responsible for continuously acquiring heterogeneous data such as power consumption, ambient temperature and humidity, and pipeline pressure. To ensure data quality, the system incorporates outlier cleaning and missing value imputation mechanisms at the edge computing gateway. Specifically, the system introduces methods such as Granger causality testing to specifically perform dual causal verification on the correlation between ambient temperature and humidity and equipment power consumption, accurately identifying channel information and time lag patterns with significant impact. This high-quality data, after deep cleaning, alignment, and the addition of causal relationship markers, is ultimately stored systematically in the factory's energy big data warehouse according to production line and workshop dimensions, providing a reliable data source for subsequent feature extraction and prediction.

[0049] Reference Figure 3 The joint extraction and coupling process of energy consumption features demonstrates the system's ability to deeply deconstruct complex electrical signals. The system first applies variational mode decomposition to the acquired raw power time-series signal, breaking it down into multiple intrinsic mode components representing normal operation, periodic fluctuations, and abnormal fluctuations. Then, multi-layer wavelet packet transform is used to further analyze the frequency domain energy distribution of each component. These multi-scale energy feature vectors are fed into a deep convolutional autoencoder, and after nonlinear dimensionality reduction mapping, low-dimensional latent vectors characterizing the equipment's operating characteristics are extracted. Simultaneously, a gated recurrent unit on the other side is specifically responsible for uncovering the dynamic coupling patterns between output and energy consumption. These two core features are finally concatenated to form a high-information-density energy consumption feature set, significantly reducing the learning difficulty of the prediction model.

[0050] Reference Figure 4 The multi-step energy consumption time-series prediction process with fusion gating demonstrates the advantages of parallel modeling of long-term and short-term features. The network architecture consists of two independent but complementary branches. One is a temporal convolutional branch composed of dilated causal convolutions, which sensitively captures short-term, drastic fluctuations in energy consumption caused by equipment start-up / shutdown or process switching through an ever-expanding receptive field. The other branch, composed of forward and backward long short-term memory units, is mainly responsible for smoothly extracting the long-term gradual trend of overall plant energy consumption. To achieve optimal fusion of the two feature streams, the system designs a learnable gating unit. This unit can adaptively calculate and allocate fusion weights for the two outputs based on the data features of the current time step. After weighted combination and mapping with a fully connected layer, it outputs a high-fidelity hourly energy demand prediction sequence.

[0051] Reference Figure 5The multi-agent collaborative energy scheduling and closed-loop optimization process demonstrates a distributed decision-making mechanism in complex factory environments. The scheduling engine treats each controllable device and energy storage system as an independent agent. During the decision-making phase, each agent not only observes its own state but also interacts with other production lines to perceive their energy usage plans and energy storage status through an attention mechanism module, effectively avoiding global load imbalance caused by information silos. Based on a near-end policy optimization algorithm, the scheduling engine aims to minimize overall cost and carbon emissions, comprehensively generating a global policy covering equipment parameter adjustment, energy storage charging and discharging, and production scheduling adjustments. After the policy is issued and executed, the actual economic and environmental benefit data generated is collected again by the system and stored in an experience replay buffer for online fine-tuning, enabling the system's scheduling capabilities to continuously evolve over long-term operation.

[0052] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart factory energy management method based on big data analytics, characterized in that, Includes the following steps: S1. Energy consumption data such as power consumption of production line equipment, temperature and humidity of workshop, compressed air pressure and three-phase electrical parameters of transformer are collected synchronously through intelligent acquisition terminals. After cleaning and interpolation by the edge gateway, an energy big data warehouse is built according to production line and workshop. S2. Perform variational mode decomposition and multi-layer wavelet packet transform on the time series data of equipment power to complete joint time-frequency feature extraction, obtain intrinsic mode components and sub-band energy input pre-trained deep convolutional autoencoder, extract the latent vector of equipment energy consumption mode, and use gated cyclic unit to extract output-energy consumption coupling features on output and energy consumption data to obtain the dynamic feature vector of energy consumption per unit product. S3. Concatenate the latent vector with the dynamic feature vector, input the pre-trained multi-step prediction network, and use a temporal convolutional gating and bidirectional long short-term memory parallel structure to extract short-term fluctuation and long-term trend features, and generate a future hourly energy consumption demand prediction sequence for each production line. S4. Input the demand forecast sequence and real-time energy price into the multi-agent reinforcement learning energy scheduling optimization engine. Use adjustable equipment power, energy storage charging and discharging power, production task scheduling deviation and demand response as actions, and use the weighted sum of energy cost and carbon emissions as the optimization objective. Use multi-agent proximal strategy optimization for offline pre-training and online fine-tuning, and output equipment parameter adjustment, energy storage plan and scheduling optimization strategy. S5. After execution, the results will be fed back to the experience playback buffer, and energy management performance reports will be generated regularly, including the energy consumption trend of each production line unit, the time-of-use energy cost composition, and the carbon emission compliance rate.

2. The intelligent factory energy management method based on big data analysis according to claim 1, characterized in that, It also includes: in step S1, the environmental temperature and humidity data of each workshop and the real-time power data of each production line equipment are verified by Granger causality test and transfer entropy analysis to complete the dual causal relationship verification, identify the significant influence channels and lag orders of environmental temperature and humidity on equipment energy consumption, and incorporate the temperature and humidity features and lag orders corresponding to the significant influence channels as exogenous inputs into the input feature set of the factory energy consumption multi-step prediction network described in step S3. In step S4, the multi-agent proximal policy optimization applies a trust domain constraint to the KL divergence of the new and old policies of each agent as a unified upper limit of the global policy update magnitude.

3. The intelligent factory energy management method based on big data analysis according to claim 1, characterized in that, It also includes: when constructing the energy scheduling optimization engine in step S4, a multi-agent cooperative communication module based on the attention mechanism is introduced. Each agent uses its own observation state as the query vector and the observation states of other agents as the key and value vectors in the decision step. The cooperative attention weight is calculated by scaling dot product attention. The weighted aggregated information of other agents is concatenated with its own observation state as the input of the policy network.

4. The intelligent factory energy management method based on big data analysis according to claim 1, characterized in that, The joint time-frequency feature extraction described in step S2 is performed as follows: variational mode decomposition is first performed on the power time series data of each production line equipment to obtain a preset number of intrinsic mode components with different center frequencies. Then, wavelet packet transform with a preset number of decomposition layers is performed on each intrinsic mode component to obtain the normalized energy distribution vector of each component in each sub-frequency band. The sub-frequency band energy distribution vectors of all intrinsic mode components are concatenated in the order of component index and then input into a deep convolutional autoencoder.

5. The intelligent factory energy management method based on big data analysis according to claim 1, characterized in that, In step S3, the temporal convolutional gating branch of the multi-step prediction network is composed of stacked causal convolutional layers with a preset number of dilated layers. The dilation factor of each layer increases exponentially with a preset base. The bidirectional long short-term memory branch is composed of forward and backward long short-term memory layers connected in parallel. The two hidden states are concatenated by time steps and mapped to the predicted values ​​through a fully connected layer. The two outputs are processed by a learnable gating fusion unit to generate fusion weights for each time step to complete adaptive weighted fusion. Then, the output is processed by a fully connected layer to generate an hourly energy consumption prediction sequence.

6. The intelligent factory energy management method based on big data analysis according to claim 1, characterized in that, The weighted combination optimization objective of energy cost and carbon emissions mentioned in step S4 is defined by the following formula: ； In the formula, Indicates that the factory contains The comprehensive optimization objective value within the scheduling cycle of each decision time step. For decision time step index, As a weighted factor for energy costs, For carbon emission weights, Assign task offset penalty weights, Indicates the first Factory energy costs at each time step Indicates the first Factory carbon emissions at each time step Indicates the first A vector of production task scheduling offsets for each time step. express Norm.

7. The intelligent factory energy management method based on big data analysis according to claim 1, characterized in that, The energy management performance report mentioned in step S5 is automatically generated at the end of each preset analysis cycle. The report includes the current cycle unit product energy consumption trend of each production line and the percentage deviation rate compared with the same period in history and industry benchmarks, the composition of the factory's total energy cost by time period and the correlation analysis of the cost ratio of each time period with the peak and valley electricity price distribution, the ratio of the factory's total carbon emissions and carbon emission compliance rate to the government-approved quota in the current cycle, and the actual adoption rate of each optimization suggestion in the scheduling strategy of the previous cycle and the energy saving quantitative assessment after adoption.

8. A smart factory energy management system based on big data analytics, characterized in that, The smart factory energy management method based on big data analysis, applied to any one of claims 1 to 7, includes the following modules: The multi-source energy data acquisition and aggregation module is equipped with a cluster of intelligent acquisition terminals distributed in various workshops of the factory and an anomaly cleaning and missing data interpolation gateway deployed at the edge. It is used to collect multi-dimensional energy consumption data across the entire factory and build a factory energy big data warehouse according to production line and workshop indexes. The joint energy consumption feature extraction module is equipped with a variational mode decomposition unit, a wavelet packet transform unit, a deep convolutional autoencoder unit, and a production value-energy consumption coupled feature extraction unit, which is used to extract low-dimensional potential vectors of energy consumption patterns and dynamic feature vectors of energy consumption per unit product from equipment power data. The energy consumption multi-step time series prediction module is equipped with a parallel temporal convolutional gated network branch, a bidirectional long short-term memory network branch, and a learnable gated fusion unit, which is used to generate an hourly energy consumption demand prediction sequence for each production line within a future preset time period. The multi-agent energy scheduling optimization module is equipped with an attention-based collaborative communication unit, a multi-agent proximal policy optimization trainer, and an experience replay buffer, which are used to output the optimal energy scheduling strategy covering equipment parameter adjustment, energy storage planning, and scheduling optimization. The scheduling execution closed-loop and performance reporting module is equipped with an actual execution result feedback unit and an automatic performance report generation unit. It is used to feed the execution results back to the optimization engine and regularly generate energy management performance reports containing energy consumption trends, cost structure and carbon emission compliance rate.

9. The intelligent factory energy management system based on big data analysis according to claim 8, characterized in that, The multi-agent energy scheduling optimization module models each adjustable device and each energy storage system as an independent reinforcement learning agent. The observation space of each agent includes the current power state of its device, the upper and lower limits of the adjustable range, and the time label of the local electricity price. The policy network adopts a three-layer fully connected hidden layer plus a Gaussian policy output layer structure. The mean of the output action is mapped to the adjustable percentage range of the rated power by tanh, and the exploration amplitude is controlled by the logarithm of the action standard deviation. The experience replay buffer of each agent is maintained independently. Global training is carried out by the central coordinator by extracting an equal number of samples from each buffer at a preset period to carry out small-batch gradient updates.

10. The intelligent factory energy management system based on big data analysis according to claim 8, characterized in that, The learnable gated fusion unit of the energy consumption multi-step time series prediction module concatenates the output features of the temporal convolutional gated branch and the output features of the bidirectional long short-term memory branch along the time step dimension. After a single-layer fully connected layer and sigmoid activation mapping, the fusion weights of the two branches at each time step are obtained. The weighted fusion feature vectors of each time step are input into the fully connected layer to generate an hourly energy consumption prediction sequence.

Images