Industrial production line energy consumption prediction system and method based on reinforcement learning and time sequence feature fusion

Abstract

The invention provides an industrial production line energy consumption prediction system and method based on reinforcement learning and time sequence feature fusion, and belongs to the technical field of industrial energy consumption prediction, and the method comprises the steps: collecting multi-source data of an industrial production line, and carrying out the preprocessing of the multi-source data to obtain a standardized time sequence data set; extracting multi-dimensional time sequence features through the fusion model, and performing dynamic weighted fusion to obtain a fusion feature vector; constructing a state space in combination with a production state identifier, inputting a prediction model based on a depth deterministic strategy gradient algorithm, and guiding training to convergence through a comprehensive reward function; real-time data is processed and then input into the trained model to obtain an energy consumption prediction result; and generating and executing a regulation and control strategy based on a prediction result, and incorporating actual operation data into the data set updating model to form closed-loop optimization. According to the invention, through multi-source data fusion, reinforcement learning model training, closed-loop optimization and exception handling, the energy consumption prediction precision, management real-time performance and operation stability of the industrial production line are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial energy consumption prediction, in particular to an industrial production line energy consumption prediction system and method based on reinforcement learning and time sequence feature fusion. BACKGROUND

[0002] In the current industrial production line energy consumption prediction technology, traditional methods mostly rely on statistical models (such as regression analysis, time series analysis) and basic machine learning algorithms (such as support vector machine, random forest), and realize prediction through fitting of historical energy consumption data. These methods can play a certain role in scenes where data rules are relatively stable, for example, short-term energy consumption estimation for single equipment or fixed production process, but they generally rely on manual feature engineering and are insufficient in correlating multi-source heterogeneous data such as equipment operation, production scheduling, and environmental changes in industrial production, making it difficult to adapt to the dynamic changes of complex industrial scenes.

[0003] With the development of industrial Internet of Things and intelligent manufacturing, energy consumption prediction technology gradually evolves towards intelligence and refinement. On the one hand, deep learning methods (such as long short-term memory network LSTM, gated recurrent unit GRU) are widely used in time series energy consumption prediction, which improves the ability to capture long-term dependencies by automatically extracting data features; on the other hand, multi-source data fusion technology has become a research hotspot, trying to integrate equipment operation parameters, production plans, environmental factors and other multi-dimensional information to improve prediction accuracy. At the same time, reinforcement learning, with its dynamic decision-making and optimization capabilities, is being explored for use in energy consumption prediction and control scenarios, aiming to achieve closed-loop management of prediction, control and feedback.

[0004] Despite the progress made by existing technologies, there are still significant limitations: first, traditional statistical models and basic machine learning methods lack the ability to model the nonlinear and strongly coupled characteristics of industrial production, and the prediction accuracy is easily affected by production condition mutations (such as equipment replacement, load adjustment); second, although a single deep learning model can handle time series data, it lacks adaptability in dynamically allocating weights to multi-source features, making it difficult to balance the influence of long-term trends and short-term fluctuations; third, most prediction methods only stop at the prediction level, without deep integration with real-time control strategies, and lack a rapid response mechanism for abnormal energy consumption deviations, limiting the practical application value of prediction results and failing to meet the high-precision and high-real-time requirements of industrial production line energy saving optimization. SUMMARY

[0005] To overcome the shortcomings of existing technologies, the present application aims to provide an industrial production line energy consumption prediction system and method based on reinforcement learning and time sequence feature fusion, which improves the energy consumption prediction accuracy, management real-time performance and operation stability of industrial production lines through multi-source data fusion, reinforcement learning model training, closed-loop optimization and abnormal handling.

[0006] To achieve the above object, the present application provides the following scheme:

[0007] An industrial production line energy consumption prediction method based on reinforcement learning and time sequence feature fusion, comprising the following steps:

[0008] S1, collecting multi-source data of the industrial production line, the multi-source data including equipment operation data, production scheduling data and environment-related data of the industrial production line, and preprocessing the same to obtain a standardized time sequence data set;

[0009] S2, based on the standardized time sequence data set, extracting multi-dimensional time sequence features through a fusion model and performing dynamic weighted fusion to obtain a fusion feature vector;

[0010] S3, based on the fusion feature vector, constructing a state space in combination with the production state identifier of the industrial production line, inputting the same into a prediction model constructed based on a deep deterministic policy gradient algorithm, and guiding training to model convergence through a comprehensive reward function to obtain a trained model capable of stably predicting the energy consumption of the industrial production line;

[0011] S4, preprocessing and feature fusion of real-time collected multi-source data of the industrial production line, generating a real-time fusion feature vector and inputting the same into the trained model to obtain an energy consumption prediction result of the industrial production line;

[0012] S5, based on the energy consumption prediction result, generating a control strategy for the industrial production line and executing the same, and incorporating actual operation data into the standardized time sequence data set, updating the trained model by regenerating a fusion feature vector to form a closed-loop optimization.

[0013] Preferably, in step S1, the equipment operation data includes current, speed and vibration data of the equipment in the industrial production line; the production scheduling data includes work order progress and material input data of the industrial production line; the environment-related data includes temperature and humidity and power grid voltage fluctuation data of the workshop where the industrial production line is located; and the preprocessing includes noise elimination, missing value processing and normalization, which are used to unify data formats and improve data quality.

[0014] Preferably, in step S2, the fusion model includes a long short-term memory network, a gated recurrent unit and an attention mechanism;

[0015] The long short-term memory network captures the long-term time sequence dependence of the energy consumption of the industrial production line through an improved forgetting gate mechanism;

[0016] The gated recurrent unit quickly responds to short-term dynamic changes in the energy consumption of the industrial production line through a simplified structure;

[0017] The attention mechanism realizes dynamic weighting through a feature weight formula, wherein the feature weight formula is:

[0018]

[0019] wherein W i is the weight of the i-th feature, S i , S j are the correlation coefficients of the i-th feature, the j-th feature and the energy consumption of the industrial production line respectively, Entropy(F i ), Entropy(F j ) are the information entropy of the i-th feature, the j-th feature respectively, and n is the total number of features.

[0020] Preferably, in step S3, the generated state identifier comprises labels of start-up, full load, model change and shutdown stages of the industrial production line; specifically comprising:

[0021] When constructing the state space, a weighted combination of the fusion feature vector and the production state identifier is adopted, and the formula is:

[0022] S = a · V + (1 - a) · L;

[0023] wherein S is the state space vector, V is the fusion feature vector, L is the production state identifier vector, and a is a dynamic adjustment parameter and Var(V) is the variance of the fusion feature vector, and Var(L) is the variance of the generated state identifier vector;

[0024] When inputting the prediction model, an action space is set based on the energy consumption regulation and control requirements of the industrial production line, the action space is a continuous energy consumption regulation and control parameter set, including the device power adjustment range and the production rhythm adjustment coefficient; the prediction model is constructed based on the deep deterministic policy gradient algorithm and adopts the Actor-Critic structure, wherein the Actor network output formula is:

[0025] μ(S|θ μ ) = σ(W2 · ReLU(W1 · S + b1) + b2);

[0026] wherein μ is the continuous energy consumption regulation and control action, θ μ is the Actor network parameter, W1 and W2 are weight matrices, b1 and b2 are bias terms, ReLU is an activation function, and σ is a mapping function;

[0027] The Critic network output formula is:

[0028] Q(S, μ|θ Q ) = W4 · ReLU(W3 · [S; μ] + b3) + b4;

[0029] wherein Q is the action value, θ QHere are the parameters of the Critic network, W3 and W4 are the weight matrices, b3 and b4 are the bias terms, and [S; μ] is the concatenated vector of the state space vector and the action space vector.

[0030] During the training process guided by the comprehensive reward function, an experience replay pool is used to store interaction samples between the state space and action space, and samples are selected based on a priority sampling mechanism according to sample value bias until the model converges; the formula for this mechanism is:

[0031]

[0032] In the formula, P k Let δ be the sampling probability of the k-th sample. k Let be the TD error of the k-th sample, e be the minimum value, and m be the total number of samples.

[0033] Preferably, in step S3, the formula for the comprehensive reward function is:

[0034]

[0035] In the formula, R t Let be the reward value at time t. For the predicted energy consumption of industrial production lines, y t Productivity(t) represents the actual energy consumption of the industrial production line, while Volatility(y) represents the unit energy output of the industrial production line at time t. t ) represents the sliding margin difference of energy consumption at time t in industrial production, and e is the minimum value to avoid the denominator being zero.

[0036] Preferably, in step S4, the energy consumption prediction result includes the energy consumption curve of the industrial production line for the next 1 to 24 hours, and the predicted value is characterized by the overconfidence interval formula, which is:

[0037]

[0038] In the formula, Z is the predicted energy consumption at time t. δ2 ε represents the quantile corresponding to the confidence level. t Let be the standard deviation of the prediction error at time t.

[0039] Preferably, in step S5, the control strategy includes equipment load allocation coefficients, process adjustment timing, and energy supply optimization thresholds for the industrial production line, which are transmitted to the control equipment of the industrial production line for execution via the industrial bus; the equipment load allocation coefficients are:

[0040]

[0041] In the formula, K iLet be the load distribution coefficient for the i-th device in the industrial production line. To maximize predicted energy consumption, Let be the predicted energy consumption of the i-th and j-th devices, respectively, and x be the total number of devices.

[0042] Preferably, in step S5, the actual operating data includes the actual energy consumption, equipment operating status, and production progress data of the industrial production line after implementing the control strategy. The standardized time-series dataset is updated with weights, and then a fused feature vector is regenerated to update the parameters of the trained model online. The formula for weight update is:

[0043] D new =β·D old +(1-β)·D real ;

[0044] In the formula, D new For the updated standardized time series dataset, D old For the original dataset, D real The data represents actual operational data, β represents the weight of historical data, and β = e -λ·Δt λ is the attenuation coefficient, and Δt is the time interval.

[0045] Preferably, it also includes an anomaly handling mechanism: first, the deviation value is calculated using the energy consumption deviation formula; when the deviation value exceeds a preset threshold, the anomaly location stage is entered; the anomaly source characteristics are located using the feature contribution formula; based on the anomaly source characteristics, the corresponding industrial production line equipment or environmental factors are associated with them, and an emergency control plan is generated; the emergency power of the equipment in the emergency control plan is calculated to ensure the stable operation of the industrial production line.

[0046] The energy consumption deviation formula is as follows:

[0047]

[0048] In the formula, Δy represents the energy consumption deviation. Let y be the predicted energy consumption at time t. t The actual energy consumption at time t;

[0049] The formula for the feature contribution is:

[0050]

[0051] In the formula, C i The contribution of the i-th feature. To predict the partial derivative of energy consumption with respect to the i-th fused feature, V i To fuse the i-th element in the feature vector, Sensitivity(F i Let ) be the sensitivity coefficient of the i-th original feature, and F i ′ represents the i-th original feature after perturbation, Var(F) i ′), Var(F i ) represent the variances of the perturbated and original features, respectively;

[0052] The formula for calculating the emergency power of the equipment in the aforementioned emergency control scheme is as follows:

[0053] P emer =P current ·(1-Δy·γ·Impact(C max ));

[0054] In the formula, P emer For the equipment's emergency power, P current Given the current power, γ is the basic emergency response factor, and C... max For the maximum feature contribution, The influence weight of the feature with the largest contribution, where n is the total number of features, and C k The contribution of the k-th feature.

[0055] This invention also provides an industrial production line energy consumption prediction system based on reinforcement learning and temporal feature fusion, used to perform the above prediction method, including:

[0056] The multi-source data acquisition module is used to collect multi-source data from industrial production lines and preprocess it to form a standardized time-series dataset.

[0057] The dynamic fusion module for temporal features is used to extract and fuse temporal features based on a standardized time-series dataset to generate a fused feature vector; the dynamic fusion module for temporal features includes a long short-term memory network, a gated recurrent unit, and an attention mechanism;

[0058] The model training module is used to construct a state space based on fused feature vectors and combined with the production status identifiers of the industrial production line. The input is a prediction model constructed based on a deep deterministic policy gradient algorithm, which is trained until convergence and then saved.

[0059] The prediction module is used to preprocess and fuse multi-source data from industrial production lines collected in real time, generate real-time fused feature vectors and input them into the trained model to obtain energy consumption prediction results.

[0060] The regulation feedback module is used to generate and execute regulation strategies based on energy consumption prediction results, incorporate actual operating data into a standardized time-series dataset, and update the trained model by regenerating fused feature vectors.

[0061] The interactive module is used for displaying relevant data from industrial production lines and providing user operation support.

[0062] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0063] (1) This invention constructs a fusion model by integrating long short-term memory network, gated recurrent unit and attention mechanism to realize dynamic weighted fusion of multi-dimensional temporal features; wherein the long short-term memory network captures long-term temporal dependencies, the gated recurrent unit responds to short-term dynamic changes, and the attention mechanism dynamically allocates weights through feature weight formula, which effectively improves the ability to represent the energy consumption related features of industrial production lines and lays the foundation for accurate prediction.

[0064] (2) This invention constructs a prediction model based on a deep deterministic strategy gradient algorithm, constructs a state space by combining production status identifiers, guides training through a comprehensive reward function, and optimizes sample selection by using a priority sampling mechanism. This enhances the model's adaptability to the nonlinear and strongly coupled characteristics of industrial production, improves the stability and accuracy of energy consumption prediction, and is especially able to cope with sudden changes in operating conditions such as equipment replacement and load adjustment.

[0065] (3) The present invention forms a closed-loop optimization mechanism of prediction, regulation and feedback. Based on the prediction results, a regulation strategy including equipment load allocation coefficient is generated, and the actual operation data is incorporated into the dataset update model. The anomaly handling mechanism ensures the stability of the production line through energy consumption deviation calculation, anomaly source location and emergency power regulation. This operation deeply integrates prediction and real-time regulation, improving the real-time performance, effectiveness and reliability of energy consumption management of industrial production lines. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a flowchart of an industrial production line energy consumption prediction method based on reinforcement learning and temporal feature fusion according to the present invention.

[0068] Figure 2 This is a schematic diagram of a module of an industrial production line energy consumption prediction system based on reinforcement learning and temporal feature fusion according to the present invention.

[0069] Figure 3 This is a comparison chart of energy consumption prediction curves of the present invention and the prior art provided in Embodiment 1 of the present invention.

[0070] Explanation of reference numerals in the attached figures:

[0071] 1. Multi-source data acquisition module; 2. Temporal feature dynamic fusion module; 3. Model training module; 4. Prediction module; 5. Regulation and feedback module; 6. Interaction module. Detailed Implementation

[0072] 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.

[0073] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0074] like Figure 1 As shown, this invention provides a method for predicting energy consumption in industrial production lines based on the fusion of reinforcement learning and temporal features, comprising the following steps:

[0075] S1. Collect multi-source data from the industrial production line, including equipment operation data, production scheduling data, and environmental correlation data of the industrial production line, and preprocess the data to obtain a standardized time-series dataset.

[0076] The equipment operation data includes current, speed, and vibration data of equipment in the industrial production line; production scheduling data includes work order progress and material input data of the industrial production line; environmental correlation data includes temperature and humidity data and power grid voltage fluctuation data of the workshop where the industrial production line is located; preprocessing includes noise removal (e.g., using wavelet transform for denoising), missing value handling (e.g., using interpolation for completion), and normalization (e.g., mapping data to the [0,1] interval) to unify the data format and improve data quality. The preprocessing methods are standard practices in existing technologies and will not be elaborated upon here; the goal is simply to achieve the purpose of unifying the data format and improving data quality.

[0077] S2. Based on the standardized time series dataset, multi-dimensional time series features are extracted through a fusion model and dynamically weighted and fused to obtain a fused feature vector.

[0078] The fusion model includes a Long Short-Term Memory (LSTM) network, a gated recurrent unit (GRU), and an attention mechanism.

[0079] LSTM captures the long-term temporal dependencies of energy consumption in industrial production lines through an improved forget gate mechanism (enhancing the impact of production status flags on forget weights);

[0080] GRU responds quickly to short-term dynamic changes in energy consumption in industrial production lines through a simplified structure (removing the cell states of LSTM and retaining only the hidden states);

[0081] The attention mechanism achieves dynamic weighting through a feature weight formula, where the feature weight formula is:

[0082]

[0083] In the formula, W i S represents the weight of the i-th feature. i S j These are the correlation coefficients between the i-th feature, the j-th feature, and the energy consumption of the industrial production line, respectively. Entropy(F) i ), Entropy(F j ) are the information entropy of the i-th feature and the j-th feature, respectively, and n is the total number of features.

[0084] S3. Based on the fused feature vector, a state space is constructed by combining the production status identifier of the industrial production line. This state space is then input into the prediction model constructed based on the deep deterministic policy gradient algorithm. The training is guided by a comprehensive reward function until the model converges, resulting in a trained model that can stably predict the energy consumption of the industrial production line.

[0085] Specifically, generating status identifiers includes labels for the start-up, full load, changeover, and shutdown stages of the industrial production line; and step S3 specifically includes the following:

[0086] When constructing the state space, a weighted combination of fused feature vectors and production state identifiers is adopted, as shown in the formula:

[0087] S = α·V + (1-α)·L;

[0088] In the formula, S is the state space vector, V is the fused feature vector, L is the production state identifier vector, and α is the dynamically adjusted parameter. Var(V) is the variance of the fused feature vector, and Var(L) is the variance of the generated state label vector;

[0089] When inputting the prediction model, an action space is set based on the energy consumption control requirements of the industrial production line. The action space is a set of continuous energy consumption control parameters, including the equipment power adjustment range and production rhythm adjustment coefficient. The prediction model is constructed based on the Deep Deterministic Policy Gradient (DDPG) algorithm and adopts an Actor-Critic structure. The Actor network output formula is:

[0090] μ(S|θ μ )=σ(W2·ReLU(W1·S+b1)+b2);

[0091] In the formula, μ represents the continuous energy consumption control action, and θ μ Here are the parameters of the Actor network, W1 and W2 are the weight matrices, b1 and b2 are the bias terms, ReLU is the activation function, and σ is the mapping function;

[0092] The Critic network output formula is:

[0093] Q(S,μ|θ Q )=W4·ReLU(W3·[S;μ]+b3)+b4;

[0094] In the formula, Q represents the action value, and θ Q Here are the parameters of the Critic network, W3 and W4 are the weight matrices, b3 and b4 are the bias terms, and [S; μ] is the concatenated vector of the state space vector and the action space vector.

[0095] During the training process guided by the comprehensive reward function, an experience replay pool is used to store interaction samples between the state space and action space, and samples are selected based on a priority sampling mechanism according to sample value bias until the model converges; the formula for this mechanism is:

[0096]

[0097] In the formula, P k Let δ be the sampling probability of the k-th sample. k Let l be the TD error of the k-th sample, l be the minimum value, taken as 0.01, and m be the total number of samples.

[0098] The formula for the above-mentioned comprehensive reward function is:

[0099]

[0100] In the formula, R t Let be the reward value at time t. For the predicted energy consumption of industrial production lines, y t Productivity(t) represents the actual energy consumption of the industrial production line, while Volatility(y) represents the unit energy output of the industrial production line at time t. t ) represents the sliding margin difference of energy consumption at time t in industrial production, and l is the minimum value to avoid the denominator being zero.

[0101] S4. Preprocess and fuse the multi-source data of the industrial production line collected in real time to generate a real-time fused feature vector and input it into the trained model to obtain the energy consumption prediction results of the industrial production line.

[0102] Specifically, using the same preprocessing method as in step S1 and the same feature fusion method as in step S2, a real-time fused feature vector is generated. This real-time fused feature vector is then input into the trained model to obtain the energy consumption prediction results for the industrial production line, including the energy consumption curves for the industrial production line over the next 1–24 hours. The predicted values ​​are represented using the overconfidence interval formula, which is:

[0103]

[0104] In the formula, Z is the predicted energy consumption at time t. δ2 ε represents the quantile corresponding to the confidence level. t Let be the standard deviation of the prediction error at time t.

[0105] S5. Based on the energy consumption prediction results, generate and execute the control strategy for the industrial production line, and incorporate the actual operating data into the standardized time series dataset. Update the trained model by regenerating the fused feature vector to form a closed-loop optimization.

[0106] Specifically, the control strategies include equipment load allocation coefficients, process adjustment timings, and energy supply optimization thresholds for the industrial production line, which are transmitted to the control equipment of the industrial production line via the industrial bus for execution; among which, the equipment load allocation coefficients are:

[0107]

[0108] In the formula, K i Let be the load distribution coefficient for the i-th device in the industrial production line. To maximize predicted energy consumption, Let be the predicted energy consumption of the i-th and j-th devices, respectively, and x be the total number of devices.

[0109] The system collects actual operational data (including actual energy consumption, equipment operating status, and production progress data) from industrial production lines after implementing control strategies. This data is then updated using a weight update formula to standardize the time-series dataset. This process regenerates the fused feature vector to update the parameters of the trained model online. The weight update formula is as follows:

[0110] D new =β·D old +(1-β)·D real ;

[0111] In the formula, D new For the updated standardized time series dataset, D old For the original dataset, D real The data represents actual operational data, β represents the weight of historical data, and β = e -λ·Δt λ is the attenuation coefficient, with a value of 0.05, and Δt is the time interval, in hours.

[0112] In addition, the above prediction method also includes an anomaly handling mechanism: first, the deviation value is calculated using the energy consumption deviation formula, and when the deviation value exceeds a preset threshold of 10%, the anomaly location stage is entered.

[0113] The energy consumption deviation formula is as follows:

[0114]

[0115] In the formula, Δy represents the energy consumption deviation. Let y be the predicted energy consumption at time t. t The actual energy consumption at time t;

[0116] The anomaly source characteristics are located using a feature contribution formula; based on these characteristics, corresponding industrial production line equipment or environmental factors are correlated to generate emergency control plans; the feature contribution formula is as follows:

[0117]

[0118] In the formula, C i The contribution of the i-th feature. To predict the partial derivative of energy consumption with respect to the i-th fused feature, V i To fuse the i-th element in the feature vector, Sensitivity(F i Let ) be the sensitivity coefficient of the i-th original feature, and F i ′ represents the i-th original feature after perturbation, Var(F) i ′), Var(F i ) represent the variances of the perturbated and original features, respectively;

[0119] The stable operation of the industrial production line is ensured by calculating the emergency power of the equipment in the emergency control plan. The formula is as follows:

[0120] P emer =P current ·(1-Δy·γ·Impact(C max ));

[0121] In the formula, P emer For the equipment's emergency power, P current Given the current power, γ is the basic emergency response factor, and C... max For the maximum feature contribution, The influence weight of the feature with the largest contribution, where n is the total number of features, and C k The contribution of the k-th feature.

[0122] Finally, this invention also provides an industrial production line energy consumption prediction system based on reinforcement learning and temporal feature fusion, used to execute the above method. For example... Figure 2 As shown, the system includes:

[0123] Multi-source data acquisition module 1 is used to collect multi-source data from industrial production lines and preprocess it to form a standardized time-series dataset.

[0124] Specifically, the multi-source data acquisition module 1 consists of sensors, PLC controllers, and environmental monitoring equipment deployed on the industrial production line. It is used to collect equipment operation data, production scheduling data, and environmental correlation data, and generates standardized time-series datasets through the built-in preprocessing unit (which realizes noise elimination, missing value processing, and normalization functions).

[0125] The dynamic fusion module 2 for temporal features is used to extract and fuse temporal features based on a standardized time-series dataset to generate a fused feature vector. The dynamic fusion module 2 for temporal features includes a long short-term memory network, a gated recurrent unit, and an attention mechanism.

[0126] Specifically, the temporal feature dynamic fusion module 2 includes an LSTM network unit, a GRU network unit, and an attention mechanism unit. The LSTM unit captures long-term temporal dependencies, the GRU unit responds to short-term dynamic changes, and the attention mechanism unit dynamically weights and fuses multi-dimensional temporal features through feature weight formulas to output a fused feature vector.

[0127] Model training module 3 is used to construct a state space based on fused feature vectors and combined with the production status identifiers of industrial production lines. The input is used to train the prediction model constructed based on the deep deterministic policy gradient algorithm until convergence and then save it.

[0128] Specifically, the model training module 3 includes a state space construction unit, a DDPG algorithm unit, and a model convergence judgment unit. The state space construction unit generates a state space by combining the fused feature vector and the production state identifier. The DDPG algorithm unit builds a prediction model based on the Actor-Critic structure and trains the model through a comprehensive reward function and a priority sampling mechanism. The model convergence judgment unit determines the model convergence and saves it by checking the change in the validation set loss (loss fluctuation less than 1% for 50 consecutive rounds).

[0129] Prediction module 4 is used to preprocess and fuse multi-source data from the industrial production line in real time, generate real-time fused feature vectors and input them into the trained model to obtain energy consumption prediction results.

[0130] Specifically, prediction module 4 receives multi-source data collected in real time, generates a real-time fused feature vector after preprocessing and feature fusion, inputs it into the trained model to output energy consumption prediction results (including 1-24h energy consumption curves and confidence intervals).

[0131] The regulation feedback module 5 is used to generate and execute regulation strategies based on energy consumption prediction results, incorporate actual operating data into a standardized time series dataset, and update the trained model by regenerating fused feature vectors.

[0132] Specifically, the control feedback module 5 includes a control strategy generation unit, an execution unit, and a model update unit; the control strategy generation unit generates strategies such as equipment load allocation and process adjustment based on the prediction results; the execution unit sends the strategies to the control equipment through the industrial bus; the model update unit collects actual operating data, updates the dataset, and fine-tunes the model parameters.

[0133] Interactive module 6 is used for displaying relevant data from the industrial production line and supporting user operations. It consists of a display screen and an operating terminal, which displays data such as energy consumption prediction curves, equipment operating status, and control strategies, and supports users in setting model parameters and viewing historical records.

[0134] The above content will be further explained below through specific implementation methods. The described embodiments are only some embodiments of the present invention, and not all embodiments.

[0135] Example 1

[0136] This embodiment takes an automotive parts production line as the application object. The production line includes 5 stamping machines and 3 assembly lines. The experimental period is 30 days. The energy consumption prediction method of this invention is compared with the traditional ARIMA time series model and the single LSTM model.

[0137] During the data acquisition phase, equipment operation data (such as current, speed, and vibration data of stamping equipment, and operating speed data of assembly lines), production scheduling data (such as work order progress and material input data), and environmental correlation data (such as workshop temperature and humidity and power grid voltage fluctuation data) of the production line are collected. Noise removal, missing value processing, and normalization are performed according to the preprocessing method described in step S1 above to obtain a standardized time-series dataset. The standardized time-series dataset is processed using a fusion model. A Long Short-Term Memory (LSTM) network captures the long-term temporal dependencies of production line energy consumption, a gated recurrent unit responds to short-term dynamic changes, and an attention mechanism dynamically weights and fuses multi-dimensional time-series features using a feature weight formula to obtain a fused feature vector. A state space is constructed by combining production status indicators such as production line start-up, full load, changeover, and shutdown. This state space is input into a prediction model built based on a deep deterministic policy gradient algorithm. Training is guided by a comprehensive reward function, and a priority sampling mechanism is used to select samples until the model converges. After preprocessing and feature fusion of the real-time multi-source data, a real-time fused feature vector is generated and input into the trained model to obtain the energy consumption prediction result. Based on the results, a control strategy is generated and implemented. Simultaneously, actual operating data is incorporated into a standardized time-series dataset, and the model is updated to form a closed-loop optimization. When energy consumption deviation exceeds a preset threshold, an anomaly handling mechanism is activated to ensure stable production line operation through energy consumption deviation calculation, anomaly source location, and emergency control.

[0138] like Figure 3 As shown, in terms of mean absolute error (MAE), the method of this invention is 2.3 kW·h, the ARIMA model is 8.7 kW·h, and the single LSTM model is 4.5 kW·h; in terms of root mean square error (RMSE), the method of this invention is 3.1 kW·h, the ARIMA model is 10.2 kW·h, and the single LSTM model is 5.8 kW·h; in terms of adaptability to sudden changes in operating conditions (error during model changeover), the method of this invention is 3.5 kW·h, the ARIMA model is 15.6 kW·h, and the single LSTM model is 7.2 kW·h. Furthermore, through the control strategy of this invention, the overall energy consumption of the production line is reduced by 8.2%, and the abnormal response time is shortened to less than 10 seconds.

[0139] The above embodiments demonstrate that the method of the present invention is significantly superior to traditional methods in terms of the accuracy of energy consumption prediction for industrial production lines and the adaptability to sudden changes in operating conditions. It can also effectively achieve dynamic optimization management of energy consumption and ensure the stable operation of the production line.

[0140] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting energy consumption in industrial production lines based on reinforcement learning and temporal feature fusion, characterized in that, Includes the following steps: S1. Collect multi-source data from the industrial production line, including equipment operation data, production scheduling data and environmental correlation data of the industrial production line, and preprocess it to obtain a standardized time-series dataset. S2. Based on the standardized time series dataset, multi-dimensional time series features are extracted through a fusion model and dynamically weighted and fused to obtain a fused feature vector; S3. Based on the fused feature vector, a state space is constructed by combining the production status identifier of the industrial production line. This state space is then input into the prediction model constructed based on the deep deterministic policy gradient algorithm. The training is guided to convergence through a comprehensive reward function, resulting in a trained model that can stably predict the energy consumption of the industrial production line. S4. Preprocess and fuse the multi-source data of the industrial production line collected in real time to generate a real-time fused feature vector and input it into the trained model to obtain the energy consumption prediction results of the industrial production line. S5. Based on the energy consumption prediction results, generate and execute the control strategy for the industrial production line, and incorporate the actual operating data into the standardized time series dataset. Update the trained model by regenerating the fused feature vector to form a closed-loop optimization.

2. The method for predicting energy consumption of industrial production lines based on reinforcement learning and temporal feature fusion as described in claim 1, characterized in that, In step S1, the equipment operation data includes the current, speed, and vibration data of the equipment in the industrial production line; the production scheduling data includes the work order progress and material input data of the industrial production line; the environmental correlation data includes the temperature and humidity of the workshop where the industrial production line is located and the power grid voltage fluctuation data; the preprocessing includes noise elimination, missing value processing, and normalization, which are used to unify the data format and improve the data quality.

3. The method for predicting energy consumption of industrial production lines based on reinforcement learning and temporal feature fusion as described in claim 1, characterized in that, In step S2, the fusion model includes a long short-term memory network, a gated recurrent unit, and an attention mechanism; The Long Short-Term Memory network captures the long-term temporal dependencies of energy consumption in industrial production lines through an improved forgetting gate mechanism. The gated circulation unit responds quickly to short-term dynamic changes in energy consumption of industrial production lines through a simplified structure; The attention mechanism achieves dynamic weighting through a feature weighting formula, wherein the feature weighting formula is: In the formula, W i S represents the weight of the i-th feature. i S j These are the correlation coefficients between the i-th feature, the j-th feature, and the energy consumption of the industrial production line, respectively. Entropy(F) i ), Entropy(F j ) are the information entropy of the i-th feature and the j-th feature, respectively, and n is the total number of features.

4. The method for predicting energy consumption of industrial production lines based on reinforcement learning and temporal feature fusion as described in claim 1, characterized in that, In step S3, the generated status identifiers include labels for the start-up, full-load, changeover, and shutdown stages of the industrial production line; specifically, they include: When constructing the state space, a weighted combination of fused feature vectors and production state identifiers is adopted, as shown in the formula: S = α·V + (1-α)·L; In the formula, S is the state space vector, V is the fused feature vector, L is the production state identifier vector, and α is the dynamically adjusted parameter. Var(V) is the variance of the fused feature vector, and Var(L) is the variance of the generated state label vector; When inputting the prediction model, an action space is set based on the energy consumption control requirements of the industrial production line. This action space is a set of continuous energy consumption control parameters, including equipment power adjustment range and production rhythm adjustment coefficients. The prediction model is constructed based on a deep deterministic strategy gradient algorithm, employing an Actor-Critic structure. The Actor network output formula is: μ(S|θ μ )=σ(W2·ReLU(W1·S+b1)+b2); In the formula, μ represents the continuous energy consumption control action, and θ μ Here are the parameters of the Actor network, W1 and W2 are the weight matrices, b1 and b2 are the bias terms, ReLU is the activation function, and σ is the mapping function; The Critic network output formula is: Q(S,μ|θ Q )=W4·ReLU(W3·[S;μ]+b3)+b4; In the formula, Q represents the action value, and θ Q Here are the parameters of the Critic network, W3 and W4 are the weight matrices, b3 and b4 are the bias terms, and [S; μ] is the concatenated vector of the state space vector and the action space vector. During the training process guided by the comprehensive reward function, an experience replay pool is used to store interaction samples between the state space and action space, and samples are selected based on a priority sampling mechanism according to sample value bias until the model converges; the formula for this mechanism is: In the formula, P k Let δ be the sampling probability of the k-th sample. k Let be the TD error of the k-th sample, e be the minimum value, and m be the total number of samples.

5. The method for predicting energy consumption of industrial production lines based on reinforcement learning and temporal feature fusion as described in claim 4, characterized in that, In step S3, the formula for the comprehensive reward function is: In the formula, R t Let be the reward value at time t. For the predicted energy consumption of industrial production lines, y t Productivity(t) represents the actual energy consumption of the industrial production line, while Volatility(y) represents the unit energy output of the industrial production line at time t. t ) represents the sliding margin difference of energy consumption at time t in industrial production, and e is the minimum value to avoid the denominator being zero.

6. The method for predicting energy consumption of industrial production lines based on reinforcement learning and temporal feature fusion as described in claim 1, characterized in that, In step S4, the energy consumption prediction result includes the energy consumption curve of the industrial production line for the next 1 to 24 hours. The predicted value is characterized by the over-confidence interval formula, which is: In the formula, Z is the predicted energy consumption at time t. δ2 ε represents the quantile corresponding to the confidence level. t Let be the standard deviation of the prediction error at time t.

7. The method for predicting energy consumption of industrial production lines based on reinforcement learning and temporal feature fusion as described in claim 1, characterized in that, In step S5, the control strategy includes the equipment load allocation coefficient, process adjustment sequence, and energy supply optimization threshold of the industrial production line, which are sent to the control equipment of the industrial production line for execution via the industrial bus; the equipment load allocation coefficient is: In the formula, K i Let be the load distribution coefficient for the i-th device in the industrial production line. To maximize predicted energy consumption, Let be the predicted energy consumption of the i-th and j-th devices, respectively, and x be the total number of devices.

8. The method for predicting energy consumption of industrial production lines based on reinforcement learning and temporal feature fusion as described in claim 7, characterized in that, In step S5, the actual operating data includes the actual energy consumption, equipment operating status, and production progress data of the industrial production line after implementing the control strategy. The standardized time-series dataset is updated with weights, and then a fused feature vector is regenerated to update the parameters of the trained model online. The formula for weight update is: D new =β·D old +(1-β)·D real ; In the formula, D new For the updated standardized time series dataset, D old For the original dataset, D real The data represents actual operational data, β represents the weight of historical data, and β = e -λ·Δt λ is the attenuation coefficient, and Δt is the time interval.

9. The method for predicting energy consumption of industrial production lines based on reinforcement learning and temporal feature fusion as described in claim 1, characterized in that, It also includes an anomaly handling mechanism: first, the deviation value is calculated using the energy consumption deviation formula; when the deviation value exceeds a preset threshold, the anomaly localization stage is entered; and the anomaly source features are located using the feature contribution formula. Based on the characteristics of the anomaly source, an emergency control plan is generated by associating it with the corresponding industrial production line equipment or environmental factors. The emergency power of the equipment in the emergency control plan is calculated to ensure the stable operation of the industrial production line. The energy consumption deviation formula is as follows: In the formula, Δy represents the energy consumption deviation. Let y be the predicted energy consumption at time t. t The actual energy consumption at time t; The formula for the feature contribution is: In the formula, C i The contribution of the i-th feature, To predict the partial derivative of energy consumption with respect to the i-th fused feature, V i To fuse the i-th element in the feature vector, Sensitivity(F i ) is the sensitivity coefficient of the i-th original feature, and F i ′ represents the i-th original feature after perturbation, Var(F) i ′), Var(F i ) represent the variances of the perturbated and original features, respectively; The formula for calculating the emergency power of the equipment in the aforementioned emergency control scheme is as follows: P emer =P current ·(1-Δy·γ·Impact(C max )); In the formula, P emer For the equipment's emergency power, P current Given the current power, γ is the basic emergency response factor, and C max For the maximum feature contribution, The influence weight of the feature with the largest contribution, where n is the total number of features, and C k The contribution of the k-th feature.

10. An industrial production line energy consumption prediction system based on reinforcement learning and temporal feature fusion, used to execute the industrial production line energy consumption prediction method based on reinforcement learning and temporal feature fusion as described in any one of claims 1 to 9, characterized in that, include: The multi-source data acquisition module is used to collect multi-source data from industrial production lines and preprocess it to form a standardized time-series dataset. The dynamic fusion module for temporal features is used to extract and fuse temporal features based on a standardized time-series dataset to generate a fused feature vector; the dynamic fusion module for temporal features includes a long short-term memory network, a gated recurrent unit, and an attention mechanism; The model training module is used to construct a state space based on fused feature vectors and combined with the production status identifiers of the industrial production line. The input is a prediction model constructed based on a deep deterministic policy gradient algorithm, which is trained until convergence and then saved. The prediction module is used to preprocess and fuse multi-source data from industrial production lines collected in real time, generate real-time fused feature vectors and input them into the trained model to obtain energy consumption prediction results. The regulation feedback module is used to generate and execute regulation strategies based on energy consumption prediction results, incorporate actual operating data into a standardized time-series dataset, and update the trained model by regenerating fused feature vectors. The interactive module is used for displaying relevant data from industrial production lines and providing user operation support.

Images