Cold and heat load prediction method applied to cigarette factory based on LSTM (Long Short Term Memory)

Abstract

The invention belongs to the technical field of cigarette factory energy supply prediction, and particularly relates to a cold and heat load prediction method applied to a cigarette factory based on LSTM (Long Short Term Memory), which comprises the steps of data collection, data preprocessing, neural network model construction, model structure definition, calculation based on an LSTM model, LSTM model training, LSTM model application and the like. Under the waste conditions of relatively independent operation of each system module, independent cold and hot energy consumption, cold and hot internal friction, waste heat emission and the like, cold and hot load curves in the next 24 hours are fitted in advance through an energy supply load prediction system, an energy flow direction trend is provided for an energy supply scheduling system, refined energy regulation and control, real-time data learning and intelligent analysis are performed, and the energy supply scheduling system is optimized. The accuracy of fault judgment of the temperature and humidity sensor by the system is improved, and the quality of cigarette production is better guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of energy supply prediction in cigarette factories, and in particular relates to a cold and hot load prediction method based on LSTM applied to cigarette factories. Background Art

[0002] In the production process of cigarette factories, both raw materials and production processes have strict requirements on the environment. The screen deviation and fluctuation of temperature and humidity will affect the quality of tobacco products, which puts higher requirements on the production environment of cigarette factories. The process production also consumes a lot of energy. The main energy media used in the process production of cigarette factories are chilled water, hot water, and steam. At present, energy is mainly supplied through fixed modes and human intervention, which has a series of problems such as extensive energy and lack of precision in energy supply.

[0003] The energy consumption of terminal equipment in cigarette factories is mainly reflected in the consumption of cold and hot loads. The traditional energy supply method is an extensive energy supply that sets a fixed output value based on experience on the energy supply side. There is often a large amount of waste heat emission, waste cold and cold and heat offset at the equipment terminal. At the same time, the production area environment of cigarette factories is complex, and multiple factors affect the operating efficiency of equipment. There are many production equipment and systems, making it difficult to effectively obtain its cold and hot load requirements and predict its cold and hot loads in each future time period. Therefore, it is impossible to optimize the adaptive energy supply and scheduling on the energy supply side. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting the heat and cold loads in a cigarette factory based on LSTM. LSTM can be used to effectively capture and process long-term dependencies in sequence data, predict the heat and cold energy load demand curve in the daily process production of the cigarette factory, and effectively avoid the gradient vanishing and gradient exploding problems that occur in traditional recurrent neural networks when processing long sequences.

[0005] To achieve the above objectives, this application is implemented through the following technical solutions:

[0006] A method for predicting cooling and heating loads in a cigarette factory based on LSTM includes the following steps:

[0007] S1. Data collection:

[0008] Collect relevant historical data and real-time data sets according to cigarette production characteristics and demand to form a time series data, and then divide the time series data into training set, validation set and test set;

[0009] S2. Data preprocessing:

[0010] Preprocess the data collected in step S1 to improve data quality by processing missing values ​​and outliers in the data, standardize and normalize the data by unifying the data format, enrich the LSTM model information by extracting feature engineering, and improve the LSTM model training effect by time series segmentation;

[0011] S3. Build a neural network model and define the model structure:

[0012] First, determine the capacity and complexity of the LSTM model. Set the number of LSTM layers and units according to the complexity of the time series features to be captured. Each LSTM unit consists of a forget gate, a memory gate, and an output gate, and contains a cell state and a hidden state.

[0013] S4. Calculation based on LSTM model:

[0014] Through a series of time steps, the gating mechanism of forget gate, input gate and output gate is used to update the cell state and hidden layer state, and finally generate a high-level feature representation of the sequence;

[0015] S5. Train the LSTM model:

[0016] The LSTM model is trained through the above steps and calculation process, and the above steps are repeated until the performance of the LSTM model on the validation set no longer improves or reaches the predetermined number of iterations. Finally, the trained LSTM model is used to predict the test set and evaluate it.

[0017] S6. LSTM model application:

[0018] The LSTM model trained and evaluated in step S5 is deployed in each subsystem that consumes cold and hot loads in process production and the supply-side output system, and the trend of the cold and hot load curve for the future set time is predicted. The predicted data is integrated and packaged and sent to the application party in the form of an interface.

[0019] Furthermore, in step S1, relevant historical data and real-time data sets are collected according to the characteristics and needs of cigarette production, including but not limited to comprehensive energy cooling consumption and heat consumption, cooling consumption and heat consumption of each subsystem, timestamp, ambient temperature and humidity, outside temperature, equipment operation quantity and efficiency, discharge volume, and one or more influencing factors of building characteristics.

[0020] Further, in step S2, the data missing value processing method includes one or more of mean filling, interpolation, forward filling or backward filling, wherein the mean filling method formula is: ,in represents the sum of all non-missing values, and n is the number of non-missing values;

[0021] Methods for processing data outliers include removal, replacement, box-and-whisker plot identification or Z-scors method.

[0022] Furthermore, in step S2, the data standardization and normalization formula is:

[0023] , where X is the original data point, is the minimum value in the data set, is the maximum value in the data set, are the normalized data points.

[0024] Furthermore, in step S2, feature engineering includes creating moving averages, differential features, or lagged features.

[0025] Furthermore, in step S3, when training the neural network model, a Dropout layer is introduced between different layers of the LSTM, that is, some neurons are randomly discarded at the output of the LSTM layer. After the last layer of the LSTM, a Dense layer is introduced to convert the output of the LSTM layer into the final prediction result. The Dense layer is used to further extract the complex relationship between features.

[0026] Furthermore, in step S4, for the basic computational flow of the LSTM unit when processing sequence data, first the forget gate decides which information to retain or discard from the cell state of the previous time step:

[0027] ,

[0028] in is the forget gate output of the current time step, is the sigmoid activation function, whose output range is between 0 and 1. is the weight matrix of the forget gate, is the hidden state of the previous time step, is the input of the current time step, is the bias term of the forget gate;

[0029] The input gate consists of two parts: the update gate and the candidate cell state:

[0030] ,

[0031] ,

[0032] in is the update gate output of the current time step, is the candidate cell state at the current time step, and They are the weight matrices for updating the gate and candidate cell states, respectively;

[0033] The cell state is updated by combining the information from the forget gate and the input gate:

[0034] ,

[0035] in is the cell state at the current time step, is the cell state at the previous time step, is element-wise multiplication;

[0036] The output gate determines what information the hidden state of the current time step should contain:

[0037] ,

[0038] ,

[0039] in is the output gate output at the current time step, is the hidden state of the current time step, is the weight matrix of the output gate.

[0040] Furthermore, while using the trained LSTM model for real-time prediction, the daily predicted values ​​and actual values ​​are used to retrain the model, making the system self-learning and adaptive, and able to continuously improve the model.

[0041] The beneficial effects of the present invention are:

[0042] During the production process of the cigarette factory, each system module operates relatively independently, and the cold and hot energy consumption is independent, resulting in waste such as cold and hot internal consumption and waste heat discharge. The parameters such as the chilled water temperature, hot water temperature and steam pressure provided by the energy supply system are provided according to the maximum guaranteed values, and it is impossible to provide accurate energy supply according to production needs, resulting in energy waste. Compared with the existing technology, this patent has the following main advantages:

[0043] (1) The energy load forecasting system can be used to fit the cold and hot load curves for the next 24 hours in advance, providing the energy supply side with control parameters for on-demand energy supply, making the output control of the energy supply system more precise and better ensuring the quality of cigarette production.

[0044] (2) The energy load forecasting system is used to fit the cold and hot load curves of each process energy system for the next 24 hours, providing the energy flow direction trend for the energy supply scheduling system, and performing refined energy regulation, real-time data learning and intelligent analysis, thereby improving the accuracy of the system's judgment on temperature and humidity sensor failures.

[0045] (3) Through machine learning and numerical empowerment, the traditional fixed-value energy supply and empiricism are replaced, thus improving the working efficiency of the energy supply system.

[0046] (4) The LSTM model has strong data processing capabilities for time series and has better nonlinear relationship processing capabilities than the ARIMA model. It has memory units and gating mechanisms, can remember important information over a long period of time, and filter out irrelevant information. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the overall framework of LSTM of the present invention.

[0048] Figure 2 This is a flow chart of the LSTM system of the present invention.

[0049] Figure 3 The figure is a schematic diagram of the LSTM system architecture and application scenarios of the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings. The following embodiments are merely exemplary and can only be used to explain and illustrate the technical solution of the present invention, but cannot be construed as limiting the technical solution of the present invention.

[0051] See also Figures 1 to 3 As shown, the present application provides a method for predicting the heat and cold loads in a cigarette factory based on LSTM, and a system for predicting the heat and cold loads in a cigarette factory based on LSTM. It proposes to use an optimized recurrent neural network, namely a long short-term memory network (LSTM), which can effectively capture and process long-term dependencies in sequence data, predict the heat and cold energy load demand curve in the daily process production of the cigarette factory, and effectively avoid the gradient vanishing and gradient exploding problems of traditional recurrent neural networks when processing long sequences.

[0052] The prediction function of the cold and hot load prediction system based on this application is mainly composed of data acquisition module, data preprocessing module and LSTM module. The data acquisition module mainly collects environmental data, production load data, time characteristics, building characteristics, equipment status and other data; the data preprocessing module mainly processes missing values ​​and outliers for the collected data to improve data quality; standardizes and normalizes data to unify data scale; uses feature engineering to extract effective features to reduce computational complexity; performs time series segmentation to improve model training effects and other processing methods. The main functions of the LSTM module are the construction of the LSTM model, model training, model evaluation and optimization, and model deployment and prediction.

[0053] S1. Data collection:

[0054] According to the production characteristics and demand of cigarettes, relevant historical data and real-time data sets are collected to form a time series data, which is then divided into training set, validation set and test set.

[0055] First, determine the data source for collection. According to the characteristics of cigarette production, the historical data set and real-time data set that need to be collected include but are not limited to comprehensive energy cooling and heat consumption, cooling and heat consumption of each subsystem, timestamp, ambient temperature and humidity, outside temperature, equipment operation quantity and efficiency, output, building characteristics and other influencing factors to form a time series of data, and then divide the data into training set, validation set, and test set.

[0056] S2. Data preprocessing:

[0057] The data collected in step S1 is preprocessed to improve data quality by processing missing values ​​and outliers in the data, standardize and normalize the data by unifying the data format, enrich the LSTM model information by extracting feature engineering, and improve the LSTM model training effect by time series segmentation.

[0058] Preprocess the data collected above to deal with missing values ​​and outliers to improve data quality; unify the data format to achieve data standardization and normalization; extract feature engineering to enrich model information; and segment the time series to improve model training effects. The main methods for dealing with missing data include mean filling, interpolation, forward filling or back-end filling, etc. Taking mean filling as an example, the formula is: ,in represents the sum of all non-missing values, and n is the number of non-missing values. The main methods for dealing with outliers in data include removal, replacement, box-and-whisker plot identification, or Z-scors. Taking the Z-scors method as an example, the formula is: , where Z is the standard score, i.e. the value of data point X after standardization, and X is the original data point, is the mean of the original data set, is the standard deviation of the original data set. >3 is considered an outlier. The standardization and normalization formula for data is: , where X is the original data point, is the minimum value in the data set, is the maximum value in the data set, is the normalized data point. The main methods for feature engineering include creating moving averages, differential features, and lag features. Taking the moving average as an example, the formula is ,in is the moving average at time point t, n is the window size of the moving average, that is, the number of consecutive data points used to calculate the average, and X is the original data value at the time point. The time series segmentation method mainly selects a fixed window size to split the time series data into multiple input-output pairs for training and testing machine learning models. The main purpose is to ensure that the data distribution between the training set and the test set is as similar as possible while maintaining the time order so that the model can learn the dynamic characteristics of the time series. In addition, the choice of segmentation method should also take into account factors such as the time dependency and seasonality of the data.

[0059] S3. Build a neural network model and define the model structure:

[0060] First, determine the capacity and complexity of the LSTM model. Set the number of LSTM layers and units according to the complexity of the time series features that need to be captured. Each LSTM unit consists of a forget gate, a memory gate, and an output gate, and contains a cell state and a hidden layer state. Specifically:

[0061] Construct a neural network model according to each system and define the model structure. First, determine the capacity and complexity of the LSTM model, and set the number of LSTM layers and units according to the complexity of the time series features required to be captured; each LSTM unit is mainly composed of a forget gate, a memory gate, and an output gate, and contains a cell state and a hidden layer state; when training the model, introduce a Dropout layer between different layers of the LSTM, that is, randomly discard some neurons at the output of the LSTM layer to prevent overfitting between layers; after the last layer of the LSTM, introduce a Dense layer to convert the output of the LSTM layer into the final prediction result; the Dense layer can further extract the complex relationship between features.

[0062] The LSTM model used mainly selectively retains or forgets information through input gate, forget gate and output gate, and uses the retained relevant information for prediction. The input gate determines which new information should be added to the memory unit, which is mainly composed of sigmoid activation function and tanh activation function, where sigmoid function determines which information is important and tanh function is used to generate new information. The output of the input gate is multiplied by the candidate information, and the result is used to update the reference of the memory unit. The forget gate determines which old information should be forgotten or removed from the memory unit, which is mainly composed of sigmoid activation function. The output of sigmoid function is directly multiplied by the current state of the memory unit to determine which information should be retained and which should be forgotten. Information with output value closer to 1 will be retained, while information with output value closer to 0 will be forgotten. The output gate determines which information in the memory unit should be output to the hidden state of the current time step, which is mainly composed of sigmoid activation function and tanh activation function, where sigmoid function determines which information should be output and tanh function is used to process the state of the memory unit to prepare for output. The output of the sigmoid function is multiplied by the state of the memory cell processed by the tanh function, and the result is the hidden state of the current time step.

[0063] S4. Calculation based on LSTM model:

[0064] The calculation based on the LSTM model mainly forgets the information in the cell state and remembers new information, transmits information useful for subsequent moment calculations, discards useless information, and outputs the hidden layer state at each time step. Through a series of time steps, the gating mechanism of the forget gate, input gate, and output gate is used to update the cell state and hidden layer state to capture the long-term dependencies in the sequence data and finally generate a high-level feature representation of the sequence.

[0065] For the basic computational flow of the LSTM unit when processing sequence data, first the forget gate decides which information to keep or discard from the cell state at the previous time step:

[0066] ,

[0067] in is the forget gate output of the current time step, is the sigmoid activation function, whose output range is between 0 and 1. is the weight matrix of the forget gate, is the hidden state of the previous time step, is the input of the current time step, is the bias term of the forget gate.

[0068] The input gate consists of two parts: the update gate and the candidate cell state. The update gate determines the update degree of the new information, and the candidate cell state generates candidates for the new state:

[0069] ,

[0070] ,

[0071] in is the update gate output of the current time step, is the candidate cell state at the current time step, and are the weight matrices for updating the gate and candidate cell states, respectively.

[0072] The cell state is updated by combining the information from the forget gate and the input gate:

[0073] ,

[0074] in is the cell state at the current time step, is the cell state at the previous time step, is element-wise multiplication.

[0075] The output gate determines what information the hidden state of the current time step should contain:

[0076] ,

[0077] ,

[0078] in is the output gate output at the current time step, is the hidden state of the current time step, is the weight matrix of the output gate. For each time step in the sequence, repeat the above steps to update the forget gate, input gate, cell state, and hidden state. The output of the LSTM layer can be the last hidden state of the sequence , or it can be a sequence of hidden states at all time steps.

[0079] The hidden state of the last time step of the LSTM layer is used as a feature vector input to the Dense layer to learn the complex mapping relationship between input and output. Define the weight matrix W and bias vector b of the Dense layer, use a linear activation function, and Perform linear transformation to get the original output of the Dense layer as the prediction result of the model.

[0080] S5. Train the LSTM model:

[0081] The LSTM model is trained through the above steps and calculation process, and the above steps are repeated until the performance of the LSTM model on the validation set no longer improves or reaches the predetermined number of iterations. Finally, the trained LSTM model is used to predict the test set and evaluate it.

[0082] Train the LSTM model through the above steps and calculation process, configure hyperparameters, set batch size and number of training rounds, and select mean square error The loss function compiles the model, uses the loss function to calculate the gradient, and updates the weight W and bias b through the back-propagation algorithm. The gradient calculation involves the chain rule, from the loss function to the activation function of the Dense layer, and then to the weight matrix and bias. Select the Adam optimizer to compile the model and use the optimization algorithm The calculated gradient is used to update the weights and biases to minimize the loss function. Repeat the above steps until the performance of the model on the validation set no longer improves or the predetermined number of iterations is reached. Finally, the trained model is used to predict the test set, and the historical data set is used as the input sequence length and the predicted cold and hot load curves for the next 24 hours are used as the output sequence length to obtain the predicted value. The mean square error is used When the RMSE is low, it means that the model prediction accuracy is high.

[0083] S6. LSTM model application:

[0084] The trained and evaluated LSTM model is deployed in each subsystem that consumes cold and hot loads in process production and the total output system on the supply side. The trend of the cold and hot load curve is predicted for the next 24 hours. The predicted data is integrated and packaged and sent to the cloud in the form of an interface. Based on the cold and hot load prediction curve data, on the energy consumption side, the energy scheduling strategy in the system is formulated and optimized in advance; on the energy supply side, the output strategy provided by the equipment system is formulated in advance, and the group control optimization strategy of the equipment, the feedforward compensation strategy of the energy supply, the energy flow distribution strategy, the energy storage strategy, etc. are optimized to improve the equipment operation efficiency and energy utilization efficiency.

[0085] By establishing LSTM models for different process production subsystems of the cigarette factory, analyzing the factors affecting the cooling and heating loads of the system operation, obtaining multi-dimensional historical data and real-time data, and processing and normalizing, the LSTM model of the subsystem is repeatedly trained until the model evaluation error is within a certain range, and then used to predict the cooling and heating load curve of the subsystem in the future, and retraining the model with the real-time data of the day, so as to achieve continuous updating and optimization of the model. Similarly, an LSTM model is established for the entire factory production system to predict the comprehensive cooling and heating load curve.

[0086] According to the cold and hot load forecast curve data of all the above subsystems, data support is provided for the formulation of energy dispatch strategies and control optimization between subsystems; according to the comprehensive cold and hot load curve data of the total system, data support is provided for the formulation of output strategies and control optimization of the equipment system on the energy supply side. Using the above data, group control optimization strategies for energy supply equipment, feedforward compensation strategies for energy supply, energy flow distribution strategies, energy storage strategies, etc. are formulated to improve equipment operation efficiency and energy utilization efficiency.

[0087] While using the trained LSTM model for real-time prediction, the daily predicted values ​​and actual values ​​are used to retrain the model, making the system self-learning and adaptive, and able to continuously improve the model.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting cooling and heating loads in a cigarette factory based on LSTM, characterized in that: The following steps are involved: S1. Data collection: Collect relevant historical data and real-time data sets according to cigarette production characteristics and demand to form a time series data, and then divide the time series data into training set, validation set and test set; S2. Data preprocessing: Preprocess the data collected in step S1 to improve data quality by processing missing values ​​and outliers in the data, standardize and normalize the data by unifying the data format, enrich the LSTM model information by extracting feature engineering, and improve the LSTM model training effect by time series segmentation; S3. Build a neural network model and define the model structure: First, determine the capacity and complexity of the LSTM model. Set the number of LSTM layers and units according to the complexity of the time series features to be captured. Each LSTM unit consists of a forget gate, a memory gate, and an output gate, and contains a cell state and a hidden state. S4. Calculation based on LSTM model: Through a series of time steps, the gating mechanism of forget gate, input gate and output gate is used to update the cell state and hidden layer state, and finally generate a high-level feature representation of the sequence; S5. Train the LSTM model: The LSTM model is trained through the above steps and calculation process, and the above steps are repeated until the performance of the LSTM model on the validation set no longer improves or reaches the predetermined number of iterations. Finally, the trained LSTM model is used to predict the test set and evaluate it. S6. LSTM model application: The LSTM model trained and evaluated in step S5 is deployed in each subsystem that consumes cold and hot loads in process production and the supply-side output system, and the trend of the cold and hot load curve for the future set time is predicted. The predicted data is integrated and packaged and sent to the application party in the form of an interface.

2. The method for predicting cooling and heating loads in a cigarette factory based on LSTM according to claim 1 is characterized in that: In step S1, relevant historical data and real-time data sets are collected according to the characteristics and needs of cigarette production, including but not limited to comprehensive energy cooling and heat consumption, cooling and heat consumption of each subsystem, timestamp, ambient temperature and humidity, outside temperature, equipment operation quantity and efficiency, discharge volume, and one or more influencing factors of building characteristics.

3. The method for predicting cooling and heating loads in a cigarette factory based on LSTM according to claim 1 is characterized in that: In step S2, the data missing value processing method includes one or more of mean filling, interpolation, forward filling or backward filling, wherein the formula of the mean filling method is: ,in represents the sum of all non-missing values, and n is the number of non-missing values; Methods for processing data outliers include removal, replacement, box-and-whisker plot identification or Z-scors method.

4. The method for predicting cooling and heating loads in a cigarette factory based on LSTM according to claim 3 is characterized in that: In step S2, the data standardization and normalization formula is: , where X is the original data point, is the minimum value in the data set, is the maximum value in the data set, are the normalized data points.

5. The method for predicting cooling and heating loads in a cigarette factory based on LSTM according to claim 3 is characterized in that: In step S2, feature engineering includes creating moving averages, differential features, or lagged features.

6. The method for predicting cooling and heating loads in a cigarette factory based on LSTM according to claim 1, characterized in that: In step S3, when training the neural network model, a Dropout layer is introduced between different layers of the LSTM, that is, some neurons are randomly discarded at the output of the LSTM layer. After the last layer of the LSTM, a Dense layer is introduced to convert the output of the LSTM layer into the final prediction result. The Dense layer is used to further extract the complex relationship between features.

7. The method for predicting cooling and heating loads in a cigarette factory based on LSTM according to claim 1, characterized in that: In step S4, for the basic computational flow of the LSTM unit when processing sequence data, first the forget gate decides which information to retain or discard from the cell state of the previous time step: ， in is the forget gate output of the current time step, is the sigmoid activation function, whose output range is between 0 and 1. is the weight matrix of the forget gate, is the hidden state of the previous time step, is the input of the current time step, is the bias term of the forget gate; The input gate consists of two parts: the update gate and the candidate cell state: ， ， in is the update gate output of the current time step, is the candidate cell state at the current time step, and They are the weight matrices for updating the gate and candidate cell states, respectively; The cell state is updated by combining the information from the forget gate and the input gate: , in is the cell state at the current time step, is the cell state at the previous time step, is element-wise multiplication; The output gate determines what information the hidden state of the current time step should contain: ， ， in is the output gate output at the current time step, is the hidden state of the current time step, is the weight matrix of the output gate.

8. The method for predicting cooling and heating loads in a cigarette factory based on LSTM according to claim 1, characterized in that: While using the trained LSTM model for real-time prediction, the daily predicted values ​​and actual values ​​are used to retrain the model, making the system self-learning and adaptive, and able to continuously improve the model.

Images