Intelligent carbon emission prediction system based on neural network

Abstract

The application discloses a carbon emission intelligent prediction system based on a neural network, and relates to the technical field of carbon emission prediction. The system acquires historical energy consumption and carbon emission data of a target enterprise, performs feature extraction and dimension reduction processing, and constructs an energy consumption dataset and a carbon emission dataset. The system captures the complex nonlinear relationship between the energy consumption data and the carbon emission data through a Siamese network, extracts potential features through a double-branch structure, and outputs a correlation matrix and nonlinear features. The system performs time series modeling on the correlation matrix and the nonlinear features through a long short-term memory network, captures the time series dependency relationship between carbon emission and energy consumption, and predicts the future carbon emission value of the target enterprise. Through a gradient descent algorithm, the system optimizes the parameter weight of the model according to the error between the predicted value and the actual value. The system provides accurate carbon emission prediction for the target enterprise and helps the enterprise make more scientific decisions in carbon emission management and energy saving and consumption reduction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of carbon emission prediction, specifically to a carbon emission intelligent prediction system based on neural networks. BACKGROUND

[0002] The carbon emission intelligent prediction system not only has important environmental significance, but also helps enterprises optimize energy management, improve resource utilization efficiency, and reduce operating costs.

[0003] For example, the Chinese invention patent with publication number CN117035167B discloses a method for predicting enterprise carbon emissions, which collects historical energy consumption data associated with enterprise carbon emission data and performs the following steps: processing historical energy consumption data, calculating the correlation degree DTW between historical energy consumption data and enterprise carbon emission data, judging the strength of the correlation between historical energy consumption data and enterprise carbon emission data, substituting strongly correlated historical energy consumption data into a long short-term memory neural network for iterative training to obtain an enterprise carbon emission prediction model, and predicting enterprise carbon emissions. The present application selects historical energy consumption data strongly correlated with enterprise carbon emissions as training data for the long short-term memory neural network through different steps, greatly reduces the correlation type, enhances the mining efficiency of the long short-term memory neural network, and makes the prediction accuracy of the obtained enterprise carbon emission prediction model higher.

[0004] However, the existing method usually selects training data by calculating the correlation degree between historical energy consumption data and carbon emission data, although it can filter out strongly correlated data, it may ignore the nonlinear relationship and complex characteristics in the data, resulting in insufficient expression ability of the model. Secondly, although the long short-term memory neural network (LSTM) can process time series data, the existing method does not fully consider the complex time sequence dependence relationship between carbon emissions and energy consumption, and cannot effectively capture long-term and short-term dependencies in time. Therefore, the existing method still needs to be improved in processing complex time series data and nonlinear relationships, and cannot accurately predict the future carbon emission value of the enterprise. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a carbon emission intelligent prediction system based on neural networks, which adopts a technical solution combining Siamese networks and long short-term memory networks, can efficiently model the complex nonlinear relationships and time sequence dependencies in historical data, thereby greatly improving the accuracy and stability of carbon emission prediction. In addition, the model parameters are adjusted by the gradient descent optimization algorithm, further improving the accuracy of the prediction results and the generalization ability of the system. The problems in the above background technology are solved.

[0006] In order to achieve the above object, the present application is realized by the following technical scheme: the carbon emission intelligent prediction system based on neural network comprises the following modules: data processing module, nonlinear relationship modeling module, prediction module and optimization module; the data processing module is used for obtaining and preprocessing the historical energy consumption data and carbon emission data of the target enterprise, including extracting key features and dimension reduction processing, and constructing energy consumption data set and carbon emission data set; the nonlinear relationship modeling module is used for capturing the complex nonlinear relationship between historical energy consumption data and carbon emission through Siamese network according to the energy consumption data set and carbon emission data set, establishing the mapping model between energy consumption data and carbon emission data, and outputting the correlation matrix and nonlinear features; the prediction module is used for time series modeling of the correlation matrix and nonlinear features through long short-term memory network, capturing the time series dependence relationship between carbon emission and energy consumption, and predicting the future carbon emission value of the target enterprise; the optimization module is used for adjusting the parameter weight of the mapping model through gradient descent algorithm according to the error between the future carbon emission value and the actual carbon emission value of the target enterprise.

[0007] Further, the specific process of obtaining and preprocessing the historical energy consumption data and historical carbon emission data of the target enterprise is as follows: obtaining the historical energy consumption data and carbon emission data of the target enterprise, cleaning the obtained data, and eliminating missing values, abnormal values and repeated data; extracting the key features of the historical energy consumption data and historical carbon emission data of the target enterprise, including energy consumption, production activity type, equipment running state and seasonal fluctuation, and performing feature dimension reduction processing.

[0008] Further, the specific process of constructing the energy consumption data set and the carbon emission data set is as follows: according to the preprocessed energy consumption data and carbon emission data of the target enterprise, a representative time window is selected for data division; the historical energy consumption data and carbon emission data are aligned according to time sequence, each time period corresponds to a group of features and target values, and the energy consumption data set and the carbon emission data set are constructed.

[0009] Further, the specific process of capturing the complex nonlinear relationship between historical energy consumption data and carbon emission through Siamese network is as follows: taking the energy consumption data set and the carbon emission data set as input, passing the data into two identical subnetworks for processing through the double-branch structure of Siamese network; in each subnetwork, the input data is extracted and embedded through the neural network layer, and the potential pattern in the historical energy consumption data and carbon emission data is extracted; the similarity of the outputs of the two subnetworks is calculated to capture the nonlinear relationship between the historical energy consumption data and carbon emission.

[0010] Further, the specific process of establishing the mapping model between energy consumption data and carbon emission data is as follows: after capturing the nonlinear relationship between historical energy consumption data and carbon emission data through the Siamese network, combine the key features of the historical data; learn the deep relationship between energy consumption and carbon emission through the fully connected layer; adjust the network weight through the regression algorithm to build the mapping model between energy consumption data and carbon emission data.

[0011] Further, the specific process of outputting the correlation matrix and nonlinear features is as follows: calculate the similarity of the outputs of the two sub-networks through the double-branch structure of the Siamese network, each sub-network processes energy consumption data and carbon emission data respectively, after extracting the features, calculate the cosine similarity to evaluate the relationship strength between each pair of data points to form a correlation matrix, each element in the matrix represents the correlation degree between the two data points, and capture the nonlinear relationship between energy consumption and carbon emission; perform nonlinear transformation on the input energy consumption data and carbon emission data through the neural network layer, in each neural network layer, process through the activation function to extract nonlinear features; output the obtained correlation matrix and nonlinear features.

[0012] Further, the specific process of capturing the time sequence dependency relationship between carbon emission and energy consumption through the long short-term memory network is as follows: organize the correlation matrix and nonlinear features into time sequence data in chronological order as the input of the long short-term memory network; use the long short-term memory network for time sequence modeling, capture the long-term and short-term dependency relationship in the historical data through the forgetting gate, input gate and output gate mechanism of the LSTM unit, and selectively retain or forget the state information at the historical time; through the multi-layer LSTM unit, gradually learn the mutual influence between energy consumption and carbon emission in different time periods, and extract the time sequence features in the data; use the feedback mechanism of LSTM to capture the time correlation of historical data based on the output of the previous time and the current input, and establish the time sequence dependency model between carbon emission and energy consumption; through the output time sequence features of the long short-term memory network, the result of capturing the time sequence dependency relationship between carbon emission and energy consumption is obtained.

[0013] Further, the specific process of predicting the future carbon emission value of the target enterprise is as follows: input the time sequence features captured through the long short-term memory network into the fully connected layer for further processing, including in the fully connected layer, fuse the time sequence features output by the multi-layer LSTM to extract the potential association information between energy consumption and carbon emission; map the time sequence features through the regression model to predict the future carbon emission value of the target enterprise.

[0014] Further, the specific process of adjusting the parameter weight of the mapping model by the gradient descent algorithm is as follows: a loss function is defined to measure the difference between the predicted value and the actual value, and the gradient of the loss function with respect to each parameter of the mapping model is calculated; a learning rate is determined to determine the step size of each parameter update, and the parameter weight of the model is updated according to the gradient and the learning rate; the gradient update step is repeated until the loss function converges.

[0015] The present application has the following advantages:

[0016] (1) The carbon emission intelligent prediction system based on neural network obtains the historical energy consumption data and carbon emission data of the target enterprise through the data processing module, and performs preprocessing, extracts key features and dimensionality reduction processing, and constructs the energy consumption data set and the carbon emission data set. This process effectively converts complex energy consumption and carbon emission data into structured data that can be used for further modeling. In addition, the nonlinear relationship modeling module captures the complex nonlinear relationship between historical energy consumption data and carbon emission through the Siamese network, establishes a mapping model between energy consumption data and carbon emission data, and outputs the correlation matrix and nonlinear features, further enhancing the system's modeling ability for the potential relationship between data.

[0017] (2) The carbon emission intelligent prediction system based on neural network uses the long short-term memory network (LSTM) to perform time series modeling on the correlation matrix and nonlinear features through the prediction module, capturing the time series dependence between carbon emission and energy consumption, thereby improving the accuracy of carbon emission prediction. The optimization module dynamically adjusts the parameter weight of the mapping model according to the error between the target enterprise's future carbon emission value and the actual carbon emission value through the gradient descent algorithm, so that the prediction model can continuously optimize and improve the accuracy and stability of the prediction. The coordinated work of these modules enables the system to accurately predict the future carbon emission value of the enterprise in different time periods, which helps enterprises to scientifically develop carbon emission reduction strategies and energy management solutions.

[0018] Of course, implementing any product of the present application does not necessarily require all the advantages described above to be achieved simultaneously. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The flowchart of the carbon emission intelligent prediction system based on neural network of the present application. DETAILED DESCRIPTION

[0020] The carbon emission intelligent prediction system based on neural network of the present application solves the problems of low accuracy of traditional carbon emission prediction methods, inability to effectively capture complex nonlinear relationships between energy consumption and carbon emission, and time series dependence.

[0021] The scheme in the present application embodiment has the following overall idea:

[0022] The historical energy consumption data and carbon emission data of the target enterprise are acquired and preprocessed, including extracting key features and dimensionality reduction processing, to construct an energy consumption dataset and a carbon emission dataset.

[0023] According to the energy consumption dataset and the carbon emission dataset, a Siamese network is used to capture the complex nonlinear relationship between the historical energy consumption data and the carbon emission, to establish a mapping model between the energy consumption data and the carbon emission data, and to output a correlation matrix and nonlinear features.

[0024] The correlation matrix and the nonlinear features are sequentially modeled by a long short-term memory network to capture the sequential dependence relationship between the carbon emission and the energy consumption, to predict the future carbon emission value of the target enterprise.

[0025] According to the error between the future carbon emission value and the actual carbon emission value of the target enterprise, the parameter weight of the mapping model is adjusted by a gradient descent algorithm.

[0026] Referring to Figure 1 The embodiment of the present application provides a technical scheme: a carbon emission intelligent prediction system based on a neural network, comprising the following modules: a data processing module, a nonlinear relationship modeling module, a prediction module, and an optimization module; the data processing module is used to acquire and preprocess the historical energy consumption data and the carbon emission data of a target enterprise, including extracting key features and dimensionality reduction processing, to construct an energy consumption dataset and a carbon emission dataset; the nonlinear relationship modeling module is used to capture the complex nonlinear relationship between the historical energy consumption data and the carbon emission according to the energy consumption dataset and the carbon emission dataset by a Siamese network, to establish a mapping model between the energy consumption data and the carbon emission data, and to output a correlation matrix and nonlinear features; the prediction module is used to sequentially model the correlation matrix and the nonlinear features by a long short-term memory network to capture the sequential dependence relationship between the carbon emission and the energy consumption, to predict the future carbon emission value of the target enterprise; and the optimization module is used to adjust the parameter weight of the mapping model by a gradient descent algorithm according to the error between the future carbon emission value and the actual carbon emission value of the target enterprise.

[0027] In this embodiment, the data processing module: the main task of this module is to obtain the historical energy consumption data and carbon emission data of the target enterprise, and to preprocess them. The preprocessing includes two aspects: extracting key features: from the original data, extract the features related to carbon emission prediction, ensure that the model can focus on the data part which has high influence on carbon emission prediction. Dimension reduction processing: in order to avoid over fitting and low computational efficiency caused by high dimension of data, dimension reduction method is used to reduce the number of data features, while retaining the key information in the data. Finally, the data processed by the data processing module will be constructed into "energy consumption dataset" and "carbon emission dataset", which will be the basis for subsequent modeling. Nonlinear relationship modeling module: the role of this module is to use Siamese network (twin neural network) to model the complex nonlinear relationship between energy consumption and carbon emission. Siamese network can efficiently learn and capture the similarity and difference between input data sets through the structure of sharing weights. The output of this module includes correlation matrix and nonlinear features, where the correlation matrix reflects the relationship strength between energy consumption data and carbon emission data, and the nonlinear features describe the complex interaction mode between them. Prediction module: the prediction module uses long short-term memory network (LSTM) to model the time series of the correlation matrix and nonlinear features output by the nonlinear relationship modeling module. LSTM is a neural network that can capture long-term and short-term dependencies in time series data, and is particularly suitable for processing time series data. Through the training of LSTM network, this module can effectively capture the time series dependence between carbon emission and energy consumption, so as to predict the future carbon emission value. The prediction result will provide quantitative basis for the future carbon emission level of the target enterprise. Optimization module: the purpose of this module is to further improve the accuracy of carbon emission prediction by optimizing the prediction result. The specific method is to calculate the error between the predicted carbon emission value and the actual carbon emission value of the target enterprise, and adjust the parameter weight of the model through gradient descent algorithm. Gradient descent algorithm will propagate the error in reverse direction to optimize the parameters of the mapping model, so that the model can more accurately predict the future carbon emission value.

[0028] Specifically, the specific process of obtaining and preprocessing the historical energy consumption data and historical carbon emission data of the target enterprise is as follows: obtaining the historical energy consumption data and carbon emission data of the target enterprise, cleaning the obtained data, eliminating missing values, outliers and duplicate data; extracting key features from the historical energy consumption data and historical carbon emission data of the target enterprise, including energy consumption, production activity type, equipment running state, seasonal fluctuation, and performing feature dimension reduction processing.

[0029] In this embodiment, the historical data of the target enterprise is obtained: first, the historical energy consumption data and carbon emission data of the target enterprise are extracted from the system or database of the target enterprise. Energy consumption data includes the consumption of various types of energy such as electricity, natural gas, steam, etc.; carbon emission data is the carbon emission calculated according to the energy consumption data. These data are usually time series, recording the energy use and carbon emission of the enterprise at different times. Data cleaning: clean the obtained data to ensure the quality and accuracy of the data. Specific operations include: removing missing values: remove records containing null or missing data to avoid the adverse effects of these missing data on subsequent analysis and modeling process. Remove outliers: identify and remove outliers that are obviously unreasonable through statistical methods or business rules. For example, some values of energy consumption or carbon emissions are much higher or lower than the reasonable range. Remove duplicate data: check the duplicate records in the data set and delete them to ensure the uniqueness of each data and avoid data redundancy affecting the analysis results. Feature extraction: extract key features from the cleaned historical data. These features include: energy consumption: the energy consumption of the target enterprise at different time periods is an important basis for predicting carbon emissions. Production activity type: different production activities have different effects on energy consumption and carbon emissions, so relevant activity type information such as production of high energy consumption products or low energy consumption products needs to be extracted. Equipment operating status: the on-off status, fault, etc. of the equipment also affects energy consumption and carbon emissions, so the operating status of the equipment needs to be recorded and extracted. Seasonal fluctuations: energy consumption and carbon emissions are usually affected by seasonal changes, especially in areas with large temperature changes, the use of heating and air conditioning equipment will cause seasonal fluctuations in energy consumption and carbon emissions, so it is very important to extract seasonal variation features. Feature dimensionality reduction: reduce the dimensionality of the extracted multiple features through technical means to reduce the data dimension and improve the training efficiency of the subsequent model. Common dimensionality reduction methods include principal component analysis (PCA), etc. By retaining the most representative features in the data, redundant information is removed, making the data more concise and more effective in supporting model training.

[0030] Specifically, the specific process of constructing the energy consumption data set and the carbon emission data set is as follows: according to the preprocessed energy consumption data and carbon emission data of the target enterprise, a representative time window is selected for data division; align the historical energy consumption data and carbon emission data according to the time sequence, each time period corresponds to a set of features and target values, and construct the energy consumption data set and the carbon emission data set.

[0031] In this embodiment, the time window is selected for data division: according to the pretreated historical energy consumption data and carbon emission data, a representative time window is selected, for example, division by month, quarter or year. These time windows help capture the periodic changes in energy consumption and carbon emissions. Data alignment: align the energy consumption data and carbon emission data in time series. That is, the energy consumption and the corresponding carbon emission in each time period should be matched to ensure that the characteristics (energy consumption data) of each time period correspond to the target value (carbon emission data). Constructing the dataset: based on the aligned data, the energy consumption data of each time period is taken as the characteristic value, and the carbon emission data is taken as the target value, and the energy consumption dataset and the carbon emission dataset are constructed respectively. In this way, each set of data contains relevant characteristics and target outputs for a time period, facilitating subsequent modeling and analysis.

[0032] Specifically, the specific process of capturing the complex nonlinear relationship between historical energy consumption data and carbon emissions through the Siamese network is as follows: the energy consumption dataset and the carbon emission dataset are taken as input, and the data is transmitted into two identical subnetworks for processing through the double-branch structure of the Siamese network; in each subnetwork, the input data is extracted and embedded through the neural network layer to extract the potential patterns in the historical energy consumption data and carbon emission data; the similarity of the outputs of the two subnetworks is calculated to capture the nonlinear relationship between the historical energy consumption data and the carbon emissions.

[0033] In this embodiment, the input data: the energy consumption dataset X energy and the carbon emission dataset X carbon represent the energy consumption and carbon emission data of the target enterprise in different time periods, respectively. Each dataset is in the form of X energy ,X carbon ∈R N ×D , where N is the number of data points and D is the feature dimension of each data point. Double-branch structure: the Siamese network adopts a double-branch structure, that is, two datasets are input into two identical subnetworks for processing. Each subnetwork has the same architecture and weights to ensure that the learned features are comparable between the two. Feature extraction and embedding: the subnetwork extracts and embeds the input data through the neural network layer. The purpose is to convert the original energy consumption data and carbon emission data into low-dimensional, representative feature vectors, usually through convolutional layers, fully connected layers, pooling layers, etc. Calculate the similarity measure: the similarity measure is calculated by calculating the similarity or distance between the feature vectors output by the two subnetworks to capture the nonlinear relationship between the data. Euclidean distance: where m is the dimension of the embedding space, and Let represent the i-th feature value. By calculating similarity or distance, the Siamese network can reveal the nonlinear relationship between historical energy consumption data and carbon emission data. During network training, the weights are optimized by minimizing a loss function (such as a contrastive loss function), which brings similar data points closer together and pushes dissimilar data points further apart, thereby capturing the complex nonlinear correlations between data.

[0034] Specifically, the process of establishing a mapping model between energy consumption data and carbon emission data is as follows: After capturing the nonlinear relationship between historical energy consumption data and carbon emission data through the Siamese network, the key features of the historical data are combined; the deep correlation between energy consumption and carbon emissions is learned through a fully connected layer; and the network weights are adjusted through a regression algorithm to construct a mapping model between energy consumption data and carbon emission data.

[0035] In this implementation scheme, the Siamese network model extracts nonlinear relationships: It captures the complex nonlinear relationships between historical energy consumption and carbon emission data using the Siamese network. The network output features (i.e., H...) carbon This includes the latent patterns and nonlinear dependencies of both datasets. Combining key features extracted from energy consumption and carbon emission data, such as energy consumption amount, production activity type, equipment operating status, and seasonal fluctuations, these features help enrich the data representation and provide more contextual information helpful for prediction. Fully connected layers learn deep associations; these key features can be further enhanced by connecting and concatenating them to the fully connected layers. Mathematical description of fully connected layers: Deep association learning is performed on features obtained from the Siamese network through fully connected layers. Fully connected layers play a crucial role at this stage, further uncovering the implicit relationship between energy consumption and carbon emissions through linear combinations and nonlinear activation functions. Application of activation functions: The input passes through a series of fully connected layers, and the output of each layer is transformed by a nonlinear activation function (such as ReLU, Sigmoid, etc.): F c =σ(W c H energy +b c ); where F c W represents the output of the c-th layer. c and b crespectively, and σ is the activation function. The regression algorithm adjusts the network weights, and by adjusting the network weights through the regression algorithm, a mapping model between energy consumption data and carbon emission data is constructed. This regression process can use the mean squared error (MSE) loss function to optimize the model, and the goal is to update the network weights by minimizing the error between the predicted carbon emission value and the actual carbon emission value. The mathematical description of the regression algorithm is to learn a suitable mapping function that can predict the corresponding carbon emission data according to the given energy consumption data. The form of the mapping model is: wherein, is the predicted carbon emission value, θ is the weight parameter of the network, H energy and are the features of the input data.

[0036] Specifically, the specific process of outputting the correlation matrix and the nonlinear features is as follows: the similarity of the outputs of the two subnetworks is calculated through the double-branch structure of the Siamese network, each subnetwork processes the energy consumption data and the carbon emission data respectively, after extracting the features, the relationship strength between each pair of data points is evaluated by calculating the cosine similarity to form the correlation matrix, each element in the matrix represents the correlation degree between the two data points, and the nonlinear relationship between energy consumption and carbon emission is captured; the input energy consumption data and carbon emission data are nonlinearly transformed through the neural network layer, in each layer of neural network, the nonlinear features are extracted after being processed by the activation function; the obtained correlation matrix and nonlinear features are taken as the output.

[0037] In this embodiment, the Siamese network is composed of two identical subnetworks, each processing energy consumption data and carbon emission data respectively. The two subnetworks share the same weights and structure, aiming to learn the similarities and differences between the two datasets. The specific process is as follows: input processing: energy consumption data and carbon emission data are input into the two subnetworks respectively. Feature extraction: each subnetwork extracts the features of the input data through multiple layers of neural networks. The output of each subnetwork contains the potential patterns and feature representations in the input data. In each subnetwork of the Siamese network, the input data (i.e. energy consumption data and carbon emission data) is processed through multiple layers of neural networks for feature extraction. The output of each layer is nonlinearly transformed by an activation function, which increases the model's nonlinear expression ability, thus capturing complex patterns in the data. In each layer of neural network, the input is weighted and summed, and the activation function converts it into more complex nonlinear features, thus better capturing the nonlinear relationship between energy consumption data and carbon emission data. Output correlation matrix and nonlinear features: correlation matrix: the correlation matrix obtained by calculating the cosine similarity, each element represents the similarity or correlation between a pair of features, and the matrix shows the mutual relationship between energy consumption and carbon emission data. Nonlinear features: the features extracted by each subnetwork after processing through neural network layers, representing complex patterns in the data, reflecting the deep nonlinear relationship of the data.

[0038] Specifically, the correlation matrix and nonlinear features are modeled by a long short-term memory network to capture the time-dependent relationship between carbon emissions and energy consumption. The specific process is as follows: organize the correlation matrix and nonlinear features in chronological order into time series data as input to the long short-term memory network; use the long short-term memory network for time series modeling, capture long-term and short-term dependencies in historical data through the forget gate, input gate and output gate mechanisms of the LSTM unit, selectively retain or forget the state information at the historical time; through multiple layers of LSTM units, gradually learn the mutual influence between energy consumption and carbon emissions at different time periods, extract time series features from the data; use the feedback mechanism of LSTM to capture the time correlation of historical data based on the output of the previous time and the current input, establish a time-dependent model between carbon emissions and energy consumption; through the output time series features of the long short-term memory network, the results of capturing the time-dependent relationship between carbon emissions and energy consumption are obtained.

[0039] In this embodiment, the correlation matrix and nonlinear features are organized into time series data in chronological order. Each time point corresponds to a data input, containing feature values related to the relationship between energy consumption and carbon emissions. These data are input into the LSTM network for further processing. The purpose of this process is to enable the network to learn the regularity of data changes over time and capture the time series dynamics between carbon emissions and energy consumption. The LSTM network captures long-term and short-term dependencies in time series data through its internal forget gate, input gate, and output gate mechanisms: Forget gate: determines how much information of the current time state should be forgotten. The forget gate generates a value between 0 and 1 based on the current input and the output of the previous time step, to determine how much historical information to forget. Input gate: determines how much new information should be stored in the memory unit of the LSTM cell at the current time. The input gate generates a value to control the amount of new information input by processing the current input and the output of the previous time step. Output gate: determines what information the LSTM cell outputs at the current time. The output gate generates an output based on the memory unit of the current state and the input signal, as the input of the next time and the final output. Capture long-term and short-term dependencies: Through the above mechanisms, LSTM can selectively retain or forget information based on long-term dependencies and short-term changes in historical data, ensuring that the model can capture important dependencies in long-term historical data while avoiding forgetting key information in the short term. This is crucial for capturing the complex time series relationship between carbon emissions and energy consumption, especially when the system is subject to periodic or sudden changes. LSTM can effectively maintain memory and make appropriate adjustments. Multi-layer LSTM learns time series features: Multi-layer LSTM units: By stacking multiple LSTM layers, the model can learn different levels of time series features in the data layer by layer. At each layer, LSTM extracts deeper time series information based on the output of its previous layer and the current input data. This multi-layer structure enables the network to learn dependencies from different time scales, better capturing the long-term and short-term effects between carbon emissions and energy consumption. Establish time series dependency model: LSTM network uses the output of the previous time and the input of the current time to capture the time correlation of historical data through its feedback mechanism. This mechanism enables LSTM to handle complex time dependencies in time series data. Through these mechanisms, LSTM gradually learns and establishes a time series dependency model between carbon emissions and energy consumption, i.e., how to predict future changes based on past energy consumption and carbon emissions data. The output of the LSTM is the time series features obtained after processing by the multi-layer neural network, which contain the time series dependency between carbon emissions and energy consumption. Through the output of the LSTM, we can get the prediction of future carbon emissions and the relationship between carbon emissions and energy consumption.

[0040] Specifically, the specific process of predicting the future carbon emission value of the target enterprise is as follows: the time series features captured by the long short-term memory network are taken as input and input to a fully connected layer for further processing. In the fully connected layer, the time series features output by multiple layers of LSTM are fused to extract potential correlation information between energy consumption and carbon emission. A regression model is used to map the time series features to predict the future carbon emission value of the target enterprise.

[0041] In the embodiment, the time series features extracted by the LSTM can be represented as a vector: wherein, is the time series feature output by the a-th LSTM unit at the t-th time point, and n is the number of LSTM layers. This vector aggregates the time series information of energy consumption and carbon emission of the target enterprise in the past period. The output of the fully connected layer is then input to a fully connected layer for further processing. The output of the fully connected layer can be represented as: t Z = σ(W·H t +b); wherein: W is the weight matrix of the fully connected layer, m is the output dimension of the fully connected layer. b is the bias term. σ is an activation function used to introduce a nonlinear transformation. The mapping of the fully connected layer output will be passed to the regression model for final prediction. Assuming that the regression model uses linear mapping, the predicted carbon emission value can be represented as: wherein: Wr is the weight matrix of the regression model, and 1 represents a single carbon emission value. br is the bias term of the regression model. is the predicted carbon emission value of the target enterprise at time t.

[0042] Specifically, the specific process of adjusting the parameter weights of the mapping model by the gradient descent algorithm is as follows: define a loss function to measure the difference between the predicted value and the actual value, calculate the gradient of the loss function with respect to each parameter of the mapping model; determine the learning rate to determine the step size of each parameter update, update the parameter weights of the model according to the gradient and the learning rate; repeat the gradient update step until the loss function converges.

[0043] In the embodiment, a loss function is defined to measure the difference between the predicted value and the actual value. In regression problems, the mean squared error (MSE) is usually used as the loss function. Given the predicted value and the actual value y t , the form of the loss function L is: wherein: T is the total number of training samples. is the predicted value of the model at time t. y tis the actual carbon emission value of the target enterprise at the moment. The gradient of the loss function with respect to each parameter of the mapping model is calculated, and the gradient is the partial derivative of the loss function with respect to each parameter, which represents the sensitivity of the loss function to the change of the parameter. Specifically, assuming that the parameters of the mapping model are and b r , we need to calculate the gradient of the loss function with respect to these parameters: the gradient of the weight: The gradient of the bias: Determine the learning rate, the learning rate η is a key hyperparameter in the gradient descent method, which determines the step size of each parameter update. A too small learning rate may cause the training process to be too slow, and a too large learning rate may cause the parameter to update too fast and miss the optimal solution. The ideal learning rate needs to be selected through tuning in practical applications. Update the parameter weight of the model, according to the gradient and the learning rate, the parameters W r and b r of the model will be updated. The update rule using gradient descent is as follows: Where: η is the learning rate. and are the gradients of the loss function with respect to W r and b r , respectively. Repeat the gradient update step until the loss function converges. The update step will be performed repeatedly, and each update will make the loss function L approach the minimum value. The convergence condition of the gradient descent algorithm can be judged according to the change of the loss function. If the change of the loss function is less than a preset threshold, or the maximum number of iterations is reached, the model is considered to have converged. Specifically, the convergence condition can be expressed as: |L( k )-L( k-1 )|<∈ where L( k ) is the loss function value of the kth iteration. L( k-1 ) is the loss function value of the previous iteration. ∈ is a very small value preset, representing the tolerance range of the convergence of the loss function.

[0044] In summary, the present application has at least the following effects:

[0045] The carbon emission intelligent prediction system based on neural network can efficiently capture the deep temporal dependence relationship between carbon emission and energy consumption by capturing the complex nonlinear relationship between historical energy consumption data and carbon emission through the Siamese network and modeling the time series characteristics through the long short-term memory network (LSTM), thereby improving the prediction accuracy. The system can perform personalized modeling and prediction according to the historical data characteristics of the target enterprise, adapt to the energy consumption mode and production activities of different enterprises, and realize customized carbon emission prediction and management scheme. The application of the Siamese network can deeply mine the complex nonlinear relationship between energy consumption data and carbon emission data, overcome the limitations of traditional linear modeling methods, and provide more accurate feature mapping for carbon emission prediction. Through the gradient descent algorithm, the mapping model parameters can be dynamically optimized to continuously improve the prediction performance of the model and ensure the high stability and low error of carbon emission prediction. The long short-term memory network (LSTM) can effectively capture the long-term and short-term dependence relationship in the historical data through its unique gating mechanism, thereby accurately modeling the time series change of carbon emission and predicting the future trend.

[0046] Those skilled in the art will appreciate that embodiments of the application can be supplied as a method, a system, or a computer program product. Accordingly, the application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage media, etc.) having computer usable program code embodied therein.

[0047] The application is described with reference to the flowcharts and / or block diagrams of the system, device (system), and computer program product according to the embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as a combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0048] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction devices that implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocksFigure 1 the function specified in the one or more blocks.

[0049] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operations steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, thus the instructions executed on the computer or other programmable data processing devices provide processes for implementing the flow Figure 1 the flow or flows and / or blocks Figure 1 the steps of the function specified in the one or more blocks.

[0050] Although preferred embodiments of the application have been described, those skilled in the art will recognize that additional modifications and variations can be made thereto without departing from the spirit and scope of the application. It is therefore intended that the appended claims cover all such modifications and variations as fall within the scope of the application.

[0051] It is apparent that a person skilled in the art can make a variety of modifications and variations to the application without departing from the spirit and scope of the application. Thus, if these modifications and variations of the application fall within the scope of the claims and their equivalents, the application is also intended to include these modifications and variations.

Claims

1. A neural network-based intelligent carbon emission prediction system, characterized in that, Comprise the following modules: data processing module, nonlinear relationship modeling module, prediction module, optimization module; The data processing module is used for obtaining historical energy consumption data and carbon emission data of a target enterprise and preprocessing, including extracting key features and dimensionality reduction processing, constructing energy consumption dataset and carbon emission dataset; The nonlinear relationship modeling module is used for capturing the complex nonlinear relationship between historical energy consumption data and carbon emission according to the energy consumption dataset and the carbon emission dataset through the Siamese network, establishing a mapping model between the energy consumption data and the carbon emission data, and outputting a correlation matrix and nonlinear features; The prediction module is used for time series modeling of the correlation matrix and the nonlinear features through the long short-term memory network, capturing the time series dependence relationship between carbon emission and energy consumption, and predicting the future carbon emission value of the target enterprise; The optimization module is used for adjusting the parameter weight of the mapping model through the gradient descent algorithm according to the error between the future carbon emission value and the actual carbon emission value of the target enterprise; The specific process of capturing the complex nonlinear relationship between historical energy consumption data and carbon emission through the Siamese network is as follows: The energy consumption dataset and the carbon emission dataset are taken as inputs, and the data is transmitted into two identical subnetworks for processing through the double-branch structure of the Siamese network; In each subnetwork, the input data is extracted and embedded through the neural network layer for feature extraction, and the potential patterns in the historical energy consumption data and the carbon emission data are extracted; The similarity of the outputs of the two subnetworks is calculated to capture the nonlinear relationship between the historical energy consumption data and the carbon emission; The specific process of establishing a mapping model between energy consumption data and carbon emission data is as follows: After capturing the nonlinear relationship between the historical energy consumption data and the carbon emission data through the Siamese network, the key features of the historical data are combined; Deep associations between energy consumption and carbon emission are learned through the fully connected layer; The mapping model between energy consumption data and carbon emission data is constructed by adjusting the network weight through the regression algorithm; The specific process of outputting the correlation matrix and the nonlinear features is as follows: The similarity of the outputs of the two subnetworks is calculated through the double-branch structure of the Siamese network, each subnetwork processes energy consumption data and carbon emission data respectively, and after feature extraction, the relationship strength between each pair of data points is evaluated by calculating the cosine similarity to form a correlation matrix, each element in the matrix represents the correlation degree between two data points, and the nonlinear relationship between energy consumption and carbon emission is captured; The input energy consumption data and carbon emission data are nonlinearly transformed through the neural network layer, and in each neural network layer, the nonlinear features are extracted through the activation function processing; The obtained correlation matrix and nonlinear features are taken as outputs. 2.The neural network-based carbon emission intelligent prediction system according to claim 1, wherein: The specific process of obtaining and preprocessing the historical energy consumption data and the historical carbon emission data of the target enterprise is as follows: The historical energy consumption data and the carbon emission data of the target enterprise are obtained, and the obtained data is cleaned to eliminate missing values, outliers and duplicate data; Extract the key features from the historical energy consumption data and historical carbon emission data of the target enterprise, including energy consumption, production activity type, equipment operation status, seasonal fluctuations, and perform feature dimensionality reduction processing. 3.The neural network-based carbon emission intelligent forecasting system according to claim 2, wherein: The specific process of constructing the energy consumption dataset and the carbon emission dataset is as follows: According to the pre-processed energy consumption data and carbon emission data of the target enterprise, select a representative time window for data division; Align the historical energy consumption data and carbon emission data according to the time sequence, each time period corresponds to a set of features and target values, construct the energy consumption dataset and the carbon emission dataset.

4. The neural network-based intelligent carbon emission prediction system according to claim 1, wherein: The specific process of capturing the time sequence dependence between carbon emissions and energy consumption through long short-term memory network is as follows: Organize the correlation matrix and nonlinear features into time series data in chronological order, as the input of long short-term memory network; Use long short-term memory network for time series modeling, through the forget gate, input gate and output gate mechanism of LSTM unit, capture the long-term and short-term dependence in historical data, selectively retain or forget the state information of historical time; Through multi-layer LSTM unit, gradually learn the mutual influence between energy consumption and carbon emission in different time periods, extract the time sequence features in the data; Use the feedback mechanism of LSTM, based on the output of the previous time and the current input, capture the time correlation of historical data, establish the time sequence dependence model between carbon emission and energy consumption; Through the output time sequence features of long short-term memory network, capture the time sequence dependence between carbon emissions and energy consumption.

5. The neural network-based intelligent carbon emission prediction system according to claim 4, wherein: The specific process of predicting the future carbon emission value of the target enterprise is as follows: Take the time sequence features captured by long short-term memory network as input, input to fully connected layer for further processing, including in the fully connected layer, fuse the time sequence features output by multi-layer LSTM, extract the potential association information between energy consumption and carbon emission; Map the time sequence features through the regression model to predict the future carbon emission value of the target enterprise. 6.The neural network-based carbon emission intelligent prediction system according to claim 5, wherein: The specific process of adjusting the parameter weights of the mapping model through gradient descent algorithm is as follows: Define the loss function to measure the difference between the predicted value and the actual value, calculate the gradient of the loss function with respect to each parameter of the mapping model; Determine the learning rate to determine the step size of each parameter update, update the parameter weights of the model according to the gradient and learning rate; Repeat the gradient update step until the loss function converges.

Images