Carbon emission calculation method based on deep learning

Abstract

The invention relates to the technical field of computers, and discloses a carbon emission calculation method based on deep learning, which comprises the following steps: S1, collecting multi-source data through a sensor, satellite remote sensing, an intelligent electric meter and a monitoring system, carrying out cleaning, denoising and missing value filling on the collected data, and in addition, carrying out data processing on the collected data; the structured data and the unstructured data are integrated through a multi-source data fusion technology; s2, performing feature extraction on the original data by adopting a deep feature learning model, and screening out key features contributing to carbon emission prediction; and S3, performing real-time decomposition and anomaly detection on the data by combining an autoregressive integral moving average model and an online learning algorithm. By combining an autoregressive integral moving average model and an online Bayesian learning method, carbon emission calculation can dynamically adjust parameters according to real-time data to adapt to environmental changes, seasonal fluctuations and emergencies, so that the flexibility and real-time performance of a prediction model are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a carbon emission calculation method based on deep learning. Background Art

[0002] Carbon emissions calculation is one of the important links in achieving the carbon neutrality goal. Existing technologies have made progress in this field. However, faced with the ever-changing environment, complex carbon emissions data and uncertain factors, existing technologies still have significant shortcomings and cannot fully meet the needs of real-time, dynamic and accurate carbon emissions calculation.

[0003] Traditional carbon emission calculation methods mostly use static models. The models cannot dynamically adjust parameters according to real-time data, and are difficult to cope with the impact of environmental changes, seasonal fluctuations and emergencies, resulting in delayed and inaccurate prediction results. The existing patent CN115934913B fills in the data gaps by matching similar units with a preset model, which solves the problem of data acquisition difficulties to a certain extent. However, its accounting process is based on static data and parameters, lacks adaptive adjustment capabilities, cannot quickly respond to dynamic changes, and is not sensitive enough to environmental fluctuations;

[0004] The existing technology handles the uncertainty in carbon emission calculations in a relatively rough way. Most methods calculate carbon emissions through average values ​​or simple estimates, lacking quantitative analysis of the uncertainty of the prediction results. The existing patent CN115934913B relies on questionnaires and similar unit data for filling, and fails to deeply quantify data uncertainty, which can easily lead to large deviations in the results, and the reliability and robustness of the prediction cannot be guaranteed;

[0005] Existing carbon emission calculation methods have significant deficiencies in feature extraction and model generalization. They usually adopt a single algorithm architecture, which makes it difficult to simultaneously capture the spatial and temporal characteristics of the data. CN115934913B performs calculations by matching data fields and filling missing data, but it lacks deep learning support in feature selection and optimization. The model architecture is single and cannot effectively deal with carbon emission prediction problems in complex data environments, resulting in limited generalization capabilities of the model.

[0006] Therefore, those skilled in the art provide a carbon emission calculation method based on deep learning to solve the problems raised in the above background technology. Summary of the invention

[0007] In view of the shortcomings of the prior art, the present invention provides a carbon emission calculation method based on deep learning to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: A carbon emission calculation method based on deep learning, comprising the following steps:

[0009] S1. Collect multi-source data through sensors, satellite remote sensing, smart meters and monitoring systems, and clean, denoise and fill missing values ​​of the collected data. In addition, integrate structured and unstructured data through multi-source data fusion technology;

[0010] S2. Use deep feature learning model to extract features from raw data and screen out key features that contribute to carbon emission prediction;

[0011] S3, by combining the autoregressive integrated moving average model and online learning algorithm, real-time data decomposition and anomaly detection;

[0012] S4. Based on the constructed feature data, a hybrid model of convolutional neural network combined with recurrent neural network or long short-term memory network is designed to predict carbon emissions;

[0013] S5. Use Bayesian deep learning algorithm combined with Markov model method to quantify the uncertainty of input data and add confidence interval to the prediction results to evaluate the reliability and risk of the prediction results;

[0014] S6. Use a dynamic factor decomposition model to adjust the carbon emission factors of different energy sources in real time, and update the emission factor formula based on actual monitoring data. The formula is as follows:

[0015] E i (t) = α i ×D i (t)+β i ×C i (t)+γ i ×M i (t),

[0016] Among them, E i (t) represents the carbon emissions of the i-th energy at time t, D i (t) is the energy consumption data, C i (t) is the environmental condition data, M i (t) is the real-time monitoring data, α i , β i , γ i is the adaptive adjustment coefficient;

[0017] S7. Combine the above models to conduct real-time prediction and analysis of carbon emissions, and output the total amount of carbon emissions and the main contributing sources;

[0018] S8. Use visualization tools to display carbon emission forecast results, trend analysis and uncertainty assessment;

[0019] S9, through the incremental learning algorithm, combined with the latest monitoring data and prediction results, the deep learning model parameters are updated in real time;

[0020] S10. Deploy the optimized deep learning model to the enterprise or power system for real-time carbon emission monitoring and intelligent scheduling.

[0021] Preferably, in the step S1, data collection includes the fusion of multi-source heterogeneous data, and specifically adopts a generative adversarial network to perform data enhancement. The algorithm formula of the generative adversarial network is as follows:

[0022] The goal of the generative adversarial network is to enable the generator to generate virtual data similar to the real data distribution through adversarial training between the generator G and the discriminator D. Its loss function is expressed as:

[0023] Among them, x~p data (x) represents the sample sampled from the real data distribution; z~p z (z) represents random noise sampled from a priori noise distribution; G(z) is a virtual sample generated by the generator; D(x) is the discriminant output probability of the discriminator for the real sample; D(G(z)) is the discriminant output probability of the discriminator for the generated sample;

[0024] The goal of the generator is to maximize the probability of the discriminator making an error. The loss function of the generator is:

[0025]

[0026] The generator optimizes its parameters by minimizing the loss to generate realistic virtual samples;

[0027] The goal of the discriminator is to maximize the ability to correctly classify real samples and generated samples, and its loss function is:

[0028]

[0029] The discriminator optimizes its parameters by minimizing the loss, thereby improving the ability to distinguish between real samples and generated samples.

[0030] Preferably, in the step S2, the autoencoder used for feature extraction is combined with a sparse regularization term during the encoding process so that more useful information is retained in the low-dimensional embedding space. The optimization formula is as follows:

[0031]

[0032] Among them, L is the loss function, X i is the input sample, To reconstruct the sample, W j is the model weight, and λ is the regularization coefficient.

[0033] Preferably, in the step S3, in the dynamic statistical analysis step, an online Bayesian learning method is used to adaptively adjust the model parameters, and the algorithm formula of the online Bayesian learning method is as follows:

[0034] Bayesian update formula: For parameter θ, the posterior probability update formula is:

[0035] Among them, P(θ|D t ) is a given data set D t The posterior distribution of P(D t |θ) is the likelihood function; P(θ|D t-1 ) is the posterior distribution of the previous moment; P(D t ) is the marginal likelihood of the data;

[0036] Definition of likelihood function: For the autoregressive integrated moving average model, assuming that the observed data y t The error is normally distributed, then the likelihood function is defined as:

[0037] Where yi is the actual observed value, is the predicted value of the autoregressive integrated moving average model, 2σ 2 is the variance of the error.

[0038] Preferably, in step S5, uncertainty quantification uses a Bayesian neural network to perform probability distribution modeling on the neural network weights to achieve uncertainty assessment of the predicted output, and the formula is:

[0039] P(Y|X)=∫P(Y|W,X)P(W|X)dW,

[0040] Among them, P(Y|X) represents the output prediction probability, and P(W|X) represents the probability distribution of the weight.

[0041] Preferably, in step S6, the adaptive adjustment coefficient α of the carbon emission factor is i , β i , Y i The objective function proxy model is established through the Bayesian optimization method to find the optimal value of the objective function. The Gaussian process is used to model the objective function, and the model is updated based on the observed data to predict the value of the unobserved point. The algorithm formula is as follows:

[0042] Gaussian process regression model:

[0043] Among them, f(x) is the objective function; f(x) is the mean function; k(x, x′) is the kernel function;

[0044] Probability distribution of the objective function: Given an observed point X = {x1, x2, ..., x n} and the corresponding target value y={y1,y2,...,y n}, the new input point x * The target value y * The predicted distribution of is:

[0045]

[0046] Where K(X, X) is the covariance matrix between observed points, and K(x*, X) is the covariance between the new observation point and the observed point.

[0047] Update of the posterior distribution: According to Bayes' theorem, the posterior mean and variance of the predicted value y* of the new point x* are: μ(x*) = K(x*, X)K(X, X)- 1 y,

[0048] σ 2 (x*)=K(x*,x*)-K(x*,X)K(X,X)- 1 K(X, x*),

[0049] Among them, μ(x*) is the predicted mean, σ 2 (x*) is the prediction variance;

[0050] Adaptive adjustment coefficient optimization: Calculate and update the adaptive adjustment coefficient α of the carbon emission factor through the Gaussian process model i , β i , γ i ,The goal is to find the optimal value of the coefficient so that the error of predicted carbon emissions is reduced;

[0051] The Bayesian optimization process determines the next set of coefficient test points based on the expected improvement or upper confidence bound strategy, as follows: EI(x) = (μ(x)-ybes t )Φ(Z)+σ(x)φ(Z),

[0052] Among them, ybes t is the optimal value in the observation, Φ(Z) is the cumulative distribution function of the standard normal distribution, and φ(Z) is the probability density function of the standard normal distribution.

[0053] Preferably, in the step S4, the carbon emission prediction model adopts a deep learning model with a hybrid architecture, the convolution layer is used to extract spatial features, the recurrent layer is used to capture time series features, and the Attention mechanism is used to optimize feature weights. The algorithm formula is as follows:

[0054] Hybrid deep learning model architecture: The model includes a combination of convolutional neural network and recurrent neural network or long short-term memory network, and uses the Attention mechanism to optimize feature weights. The model architecture is as follows:

[0055] Input layer: Input carbon emission related multidimensional time series data X = {x1, x2, ..., x T}, where each x t is the eigenvector at time t;

[0056] Convolutional layer: used to extract local spatial features in the data, and perform feature extraction through multiple convolution kernels: Among them, W i is the i-th convolution kernel, * indicates the convolution operation, b i is the bias term;

[0057] Recurrent layer: used to capture long-term dependencies in time series. The core of the recurrent layer is the cell state and gating mechanism.

[0058] The update formula of the loop layer is:

[0059] The information that the forget gate determines to be retained: f t =σ(W f ·[h t-1 , x t ]+b f ),

[0060] The input gate determines the new information to be updated: i t =σ(W i ·[h t-1 , x t ]+b i ),

[0061]

[0062] Cell status update:

[0063] The output gate determines the amount of information output: o t =σ(W o ·[h t-1 , x t ]+b o ),

[0064] h t =o t *tanh(C t ).

[0065] Preferably, the Attention mechanism is used to assign different importance weights to input features to optimize the utilization of features. The Attention score calculation formula is:

[0066] e t =v T tanh(W h h t +W s s t-1 ),

[0067] Among them, e t is the attention score; W h , W s is the weight matrix; s t-1 is the hidden state of the previous time step; v is the vector used to generate attention weights;

[0068] Attention weight calculation:

[0069] Context vector generation:

[0070] Context vector c t The hidden states at different time steps are weighted and summed to reflect the importance;

[0071] Output layer: Use context vector c t Generate the final carbon emissions forecast:

[0072]

[0073] Among them, W y is the output layer weight matrix, b y is the output bias.

[0074] Preferably, in step S5, a hidden Markov model is used to capture the hidden state transition characteristics in the time series, and the HMM consists of a hidden state, an observed state, a state transition probability, an observed probability, and an initial probability distribution; the basic elements of the hidden Markov model are:

[0075] S = {s1, s2, ..., s N}: hidden state collection;

[0076] O = {o1, o2, ..., o T}: observation sequence;

[0077] A=[a ij ]: state transition probability matrix, a ij =P(s j |s i );

[0078] B=[b j (o)]: observation probability matrix, b j (o) = P(o|sj );

[0079] π=[π i ]: initial state probability distribution, π i =P(s i );

[0080] Forward algorithm: used to calculate the probability P(O|λ) of a given observation sequence;

[0081] Define the forward variable α t (i) indicates that the state at time t is s i And we observe o1, o2, ..., o t Probability: α t (i) = P(o1, o2, ..., o t ,s i |λ),

[0082] The recursive formula is:

[0083] Initialization: α1(i) = π i b i (o1),

[0084] Recursion:

[0085] termination:

[0086] Preferably, in step S8, the result visualization module can update the key indicators of carbon emission prediction in real time, display uncertainty analysis, trend prediction and sources of abnormal fluctuations, and provide intelligent support for management decisions.

[0087] The present invention provides a carbon emission calculation method based on deep learning. It has the following beneficial effects:

[0088] 1. The present invention combines the autoregressive integrated moving average model and the online Bayesian learning method, so that the carbon emission calculation can dynamically adjust parameters according to real-time data to adapt to environmental changes, seasonal fluctuations and emergencies, so as to significantly improve the flexibility and real-time performance of the prediction model. This is an innovative breakthrough that cannot be achieved by traditional static models, making carbon emission management more accurate and timely.

[0089] 2. The present invention quantifies the uncertainty of the prediction results through Bayesian deep learning and Gaussian process regression, provides confidence intervals and risk assessments for carbon emission predictions, effectively reduces the uncertainty in decision-making, and provides reliability assessments of prediction results. It is a highly innovative improvement in carbon emission calculations and provides more robust support for management decisions.

[0090] 3. The present invention adopts a hybrid architecture of convolutional neural network and recurrent neural network, and introduces the Attention mechanism to optimize feature weights, which can fully capture the spatial and temporal characteristics of the data, significantly improve the accuracy of carbon emission prediction and the generalization ability of the model. Its innovative architecture breaks through the limitations of the traditional single network, making the performance of carbon emission prediction in complex environments even better, demonstrating the advanced application of deep learning in the field of environmental science. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0092] In order to make the technical personnel in the technical field understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in combination with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a partial embodiment of the present invention, not a complete embodiment. Based on the embodiment of the present invention, other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0093] The present invention is described in detail below in conjunction with the accompanying drawings:

[0094] Example:

[0095] Please see attached Figure 1 , an embodiment of the present invention provides a carbon emission calculation method based on deep learning, comprising the following steps: S1, collecting multi-source data through sensors, satellite remote sensing, smart meters and monitoring systems, and cleaning, denoising and filling missing values ​​of the collected data, and integrating structured and unstructured data through multi-source data fusion technology;

[0096] S2. Use deep feature learning model to extract features from raw data and screen out key features that contribute to carbon emission prediction;

[0097] S3, by combining the autoregressive integrated moving average model and online learning algorithm, real-time data decomposition and anomaly detection;

[0098] S4. Based on the constructed feature data, a hybrid model of convolutional neural network combined with recurrent neural network or long short-term memory network is designed to predict carbon emissions;

[0099] S5. Use Bayesian deep learning algorithm combined with Markov model method to quantify the uncertainty of input data and add confidence interval to the prediction results to evaluate the reliability and risk of the prediction results;

[0100] S6. Use a dynamic factor decomposition model to adjust the carbon emission factors of different energy sources in real time, and update the emission factor formula based on actual monitoring data. The formula is as follows:

[0101] E i (t) = α i ×D i (t)+β i ×C i (t)+γ i ×M i (t),

[0102] Among them, E i (t) represents the carbon emissions of the i-th energy at time t, D i (t) is the energy consumption data, C i (t) is the environmental condition data, M i (t) is the real-time monitoring data, α i , β i , γ i is the adaptive adjustment coefficient;

[0103] S7. Combine the above models to conduct real-time prediction and analysis of carbon emissions, and output the total amount of carbon emissions and the main contributing sources;

[0104] S8. Use visualization tools to display carbon emission forecast results, trend analysis and uncertainty assessment;

[0105] S9, through the incremental learning algorithm, combined with the latest monitoring data and prediction results, the deep learning model parameters are updated in real time;

[0106] S10. Deploy the optimized deep learning model to the enterprise or power system for real-time carbon emission monitoring and intelligent scheduling.

[0107] The benefit of S1 is that it enriches data dimensions, improves the comprehensiveness of predictions, and effectively integrates structured and unstructured data; the benefit of S2 is that it reduces the workload of manual feature engineering, improves the efficiency and accuracy of feature extraction, reduces the interference of noise features, and significantly improves the prediction effect of the model; the benefit of S3 is that it can quickly identify abnormal fluctuations in the data, provide early warnings for timely response to abnormal changes in carbon emissions, avoid erroneous predictions, and improve the flexibility of the model; the benefit of S4 is that it can utilize spatial and temporal characteristics to effectively improve the accuracy of carbon emission predictions, which helps to make more accurate emission predictions; the benefit of S5 is that it can provide decision makers with information on the credibility of prediction results, which helps to formulate carbon emission control policies more robustly. The benefit of S6 is to improve the accuracy of emission estimates and avoid the deviation caused by the use of static factors; the benefit of S7 is to support the immediate decision-making of enterprises and systems, provide a basis for the formulation of accurate emission reduction strategies, and make regulatory measures faster and more accurate; the benefit of S8 is to facilitate users at different levels to understand and use data, help to detect potential problems and evaluate policy effects, and provide managers with a clear decision support interface; the benefit of S9 is to avoid prediction failures caused by environmental changes, save computing resources, and improve system operation efficiency; the benefit of S10 is to provide real-time carbon emission information, support dynamic scheduling and management, optimize energy use strategies, reduce carbon footprint, and achieve green development.

[0108] In step S1, data collection includes the fusion of multi-source heterogeneous data. Specifically, a generative adversarial network is used for data enhancement. The algorithm formula of the generative adversarial network is as follows:

[0109] The goal of the generative adversarial network is to enable the generator to generate virtual data similar to the real data distribution through adversarial training between the generator G and the discriminator D. Its loss function is expressed as:

[0110]

[0111] Among them, x~p data (x) represents the sample sampled from the real data distribution; z~p z (z) represents random noise sampled from a priori noise distribution; G(z) is a virtual sample generated by the generator; D(x) is the discriminant output probability of the discriminator for the real sample; D(G(z)) is the discriminant output probability of the discriminator for the generated sample;

[0112] The goal of the generator is to maximize the probability of the discriminator making an error. The loss function of the generator is:

[0113]

[0114] The generator optimizes its parameters by minimizing the loss to generate realistic virtual samples;

[0115] The goal of the discriminator is to maximize the ability to correctly classify real samples and generated samples, and its loss function is:

[0116] The discriminator optimizes its parameters by minimizing the loss, thereby improving the ability to distinguish between real samples and generated samples.

[0117] Through adversarial training of the generator and the discriminator, the generative adversarial network can continuously optimize the quality and authenticity of the generated data. Its mechanism makes the generative adversarial network particularly suitable for data enhancement and multi-source data fusion. In particular, in carbon emission calculations, by generating high-quality virtual data, it can effectively supplement training data and improve the prediction performance of the model.

[0118] The algorithm generates realistic data samples, improves data diversity and representativeness, overcomes the problem of data scarcity, can effectively handle the fusion of multi-source heterogeneous data, and enhance the adaptability of deep learning models to complex scenarios.

[0119] In step S2, the autoencoder used for feature extraction combines sparse regularization terms in the encoding process so that more useful information is retained in the low-dimensional embedding space. The optimization formula is as follows:

[0120]

[0121] Among them, L is the loss function, X i is the input sample, To reconstruct the sample, W j is the model weight, and λ is the regularization coefficient.

[0122] The autoencoder combined with the optimization formula of the sparse regularization term can effectively map the input data to a low-dimensional space, retain key information, reduce noise and redundancy, and improve the quality of feature expression;

[0123] Through sparse regularization, the model can suppress the influence of useless features, reduce the risk of overfitting, and maintain good performance on new data sets;

[0124] The sparse characteristics of the autoencoder help to effectively compress high-dimensional data, making the low-dimensional embedding space more streamlined, which is of great significance for feature selection in data analysis and carbon emission calculations.

[0125] In step S3, in the dynamic statistical analysis step, the online Bayesian learning method is used to adaptively adjust the model parameters. The algorithm formula of the online Bayesian learning method is as follows:

[0126] Bayesian update formula: For parameter θ, the posterior probability update formula is:

[0127] Among them, P(θ|D t ) is a given data set D t The posterior distribution of P(D t |θ) is the likelihood function; P(θ|D t-1 ) is the posterior distribution of the previous moment; P(D t ) is the marginal likelihood of the data;

[0128] Definition of likelihood function: For the autoregressive integrated moving average model, assuming that the observed data y t The error is normally distributed, then the likelihood function is defined as:

[0129] Where yi is the actual observed value, is the predicted value of the autoregressive integrated moving average model, 2σ 2 is the variance of the error.

[0130] Through the Bayesian update formula and likelihood function, the model can continuously make adaptive adjustments to ensure that parameters are continuously optimized and adapt to the ever-changing data environment. Especially in carbon emission calculations, emission data and influencing factors will change dynamically over time. Compared with traditional batch learning methods, online Bayesian learning methods can gradually update parameters, avoid frequent retraining, significantly reduce computing costs, and improve efficiency.

[0131] By quantifying parameter uncertainty and continuously updating it, online Bayesian learning methods make model predictions more reliable and can provide robust decision support in the face of uncertainty and data fluctuations.

[0132] In step S5, uncertainty quantification uses a Bayesian neural network to model the probability distribution of the neural network weights and implement uncertainty assessment of the predicted output. The formula is:

[0133] P(Y|X)=∫P(Y|W,X)P(W|X)dW,

[0134] Among them, P(Y|X) represents the output prediction probability, and P(W|X) represents the probability distribution of the weight.

[0135] The Bayesian neural network gives a prediction value and can also give a range of uncertainty in the prediction, making the model's output more interpretable and credible. By dynamically adjusting the weight distribution, the Bayesian neural network can continuously update the prediction based on the arrival of new data and adapt to environmental changes, helping to improve the accuracy of carbon emissions calculations. Its uncertainty quantification method greatly enhances the model's ability to handle complex problems, especially when faced with high noise, missing data or new environments. The Bayesian neural network performs more robustly.

[0136] In step S6, the adaptive adjustment coefficient α of the carbon emission factor i, β i , γ i The objective function proxy model is established through the Bayesian optimization method to find the optimal value of the objective function. The Gaussian process is used to model the objective function, and the model is updated based on the observed data to predict the value of the unobserved point. The algorithm formula is as follows:

[0137] Gaussian process regression model:

[0138] Among them, f(x) is the objective function; f(x) is the mean function; k(x, x′) is the kernel function;

[0139] Probability distribution of the objective function: Given an observed point X = {x1, x2, ..., x n} and the corresponding target value y={y1,y2,...,y n}, the new input point x * The target value y * The predicted distribution of is:

[0140]

[0141] Where K(X, X) is the covariance matrix between observed points, and K(x*, X) is the covariance between the new observation point and the observed point.

[0142] Update of the posterior distribution: According to Bayes' theorem, the posterior mean and variance of the predicted value y* of the new point x* are: μ(x * )=K(x * ,X)K(X,X) -1 y,

[0143] σ 2 (x * )=K(x * , x * )-K(x * ,X)K(X,X) -1 K(X,x * ),

[0144] Among them, μ(x * ) is the predicted mean, σ 2 (x * ) is the prediction variance;

[0145] Adaptive adjustment coefficient optimization: Calculate and update the adaptive adjustment coefficient α of the carbon emission factor through the Gaussian process model i , β i , γ i ,The goal is to find the optimal value of the coefficient so that the error of predicted carbon emissions is reduced;

[0146] The Bayesian optimization process determines the next set of test points for the coefficients based on the expected improvement or upper confidence bound strategy, as follows: EI(x) = (μ(x) - y best )Φ(Z)+σ(x)φ(Z),

[0147] Among them, y best is the optimal value in the observation, Φ(Z) is the cumulative distribution function of the standard normal distribution, and φ(Z) is the probability density function of the standard normal distribution.

[0148] Gaussian process combined with Bayesian optimization can provide an accurate and efficient optimization method when faced with complex, multivariable carbon emission factor adjustments; through real-time updates of predicted means and variances, the optimization process can identify high-potential parameter combinations, while quantifying risks to ensure that the adjustment process is robust and reliable. It is expected that improved strategies will help the model quickly determine the optimal solution in an uncertain environment, providing an accurate adjustment basis for carbon emission control and energy management.

[0149] In step S4, the carbon emission prediction model adopts a deep learning model with a hybrid architecture. The convolution layer is used to extract spatial features, the recurrent layer is used to capture time series features, and the Attention mechanism is used to optimize feature weights. The algorithm formula is as follows:

[0150] Hybrid deep learning model architecture: The model includes a combination of convolutional neural network and recurrent neural network or long short-term memory network, and uses the Attention mechanism to optimize feature weights. The model architecture is as follows:

[0151] Input layer: Input carbon emission related multidimensional time series data X = {x1, x2, ..., x T}, where each x t is the eigenvector at time t;

[0152] Convolutional layer: used to extract local spatial features in the data, and perform feature extraction through multiple convolution kernels: Among them, W i is the i-th convolution kernel, * indicates the convolution operation, b i is the bias term;

[0153] Recurrent layer: used to capture long-term dependencies in time series. The core of the recurrent layer is the cell state and gating mechanism.

[0154] The update formula of the loop layer is:

[0155] The information that the forget gate determines to be retained: f t =σ(W f ·[h t-1 , x t ]+b f ),

[0156] The input gate determines the new information to be updated: i t =σ(W i ·[h t-1 , x t ]+b i ),

[0157]

[0158] Cell status update:

[0159] The output gate determines the amount of information output: o t =σ(W o ·[h t-1 , x t ]+b o ),

[0160] h t =o t *tanh(C t );

[0161] The Attention mechanism is used to assign different importance weights to input features to optimize feature utilization. The Attention score calculation formula is:

[0162] e t =v T tanh(W h h t +W s s t-1 ),

[0163] Among them, e t is the attention score; W h , W s is the weight matrix; s t-1 is the hidden state of the previous time step; v is the vector used to generate attention weights;

[0164] Attention weight calculation:

[0165] Context vector generation:

[0166] Context vector c t The hidden states at different time steps are weighted and summed to reflect the importance;

[0167] Output layer: Use context vector c t Generate the final carbon emissions forecast:

[0168]

[0169] Among them, W y is the output layer weight matrix, b y is the output bias.

[0170] The convolution layer extracts spatial features, the recurrent layer captures temporal dependencies, and the Attention mechanism further optimizes feature weights, allowing the model to simultaneously process complex multi-dimensional time series data. Through the Attention mechanism, the model improves prediction accuracy and enhances understanding of the key role of features, providing explanatory support for carbon emission management. Through the synergy of each layer, the model can achieve efficient computing, real-time adjustment, and accurate prediction in complex carbon emission prediction tasks, providing strong technical support for emission reduction decisions.

[0171] In step S5, the hidden Markov model is used to capture the hidden state transition characteristics in the time series. The HMM consists of hidden state, observed state, state transition probability, observation probability and initial probability distribution; the basic elements of the hidden Markov model are:

[0172] S = {s1, s2, ..., s N}: hidden state collection;

[0173] O = {o1, o2, ..., o T}: observation sequence;

[0174] A=[a ij ]: state transition probability matrix, a ij =P(s j |s i );

[0175] B=[b j (o)]: observation probability matrix, b j (o) = P(o|s j );

[0176] π=[π i ]: initial state probability distribution, π i =P(s i );

[0177] Forward algorithm: used to calculate the probability P(O|λ) of a given observation sequence;

[0178] Define the forward variable α t (i) indicates that the state at time t is s i And we observe o1, o2, ..., o t Probability: α t (i) = P(o1, o2, ..., o t ,s i |λ),

[0179] The recursive formula is:

[0180] Initialization: α1(i) = π i b i (o1),

[0181] Recursion:

[0182] termination:

[0183] The hidden Markov model links hidden states and observed data through a forward algorithm, revealing the implicit dynamic patterns in time series and providing a tool for understanding the potential driving factors of carbon emission changes. The hidden Markov model can capture the transition characteristics in time series, adapt to changes in different time steps, and provide data support for predicting future states and carbon emission levels. By inferring hidden states, the hidden Markov model can explain the fluctuations and trends in carbon emission sequences and provide a more accurate basis for decision-making and regulation.

[0184] In step S8, the result visualization module can update the key indicators of carbon emission forecasting in real time, display uncertainty analysis, trend forecasting and sources of abnormal fluctuations, and provide intelligent support for management decisions.

[0185] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A carbon emission calculation method based on deep learning, characterized in that: The steps include: S1. Collect multi-source data through sensors, satellite remote sensing, smart meters and monitoring systems, and clean, denoise and fill missing values ​​of the collected data. In addition, integrate structured and unstructured data through multi-source data fusion technology; S2. Use deep feature learning model to extract features from raw data and screen out key features that contribute to carbon emission prediction; S3, by combining the autoregressive integrated moving average model and online learning algorithm, real-time data decomposition and anomaly detection; S4. Based on the constructed feature data, a hybrid model of convolutional neural network combined with recurrent neural network or long short-term memory network is designed to predict carbon emissions; S5. Use Bayesian deep learning algorithm combined with Markov model method to quantify the uncertainty of input data and add confidence interval to the prediction results to evaluate the reliability and risk of the prediction results; S6. Use a dynamic factor decomposition model to adjust the carbon emission factors of different energy sources in real time, and update the emission factor formula based on actual monitoring data. The formula is as follows: E i (t)=a i ×D i (t)+β i ×C i (t)+γ i ×M i (t), Among them, E i (t) represents the carbon emissions of the i-th energy at time t, D i (t) is the energy consumption data, C i (t) is the environmental condition data, M i (t) is the real-time monitoring data, α i , β i , γ i is the adaptive adjustment coefficient; S7. Combine the above models to conduct real-time prediction and analysis of carbon emissions, and output the total amount of carbon emissions and the main contributing sources; S8. Use visualization tools to display carbon emission forecast results, trend analysis and uncertainty assessment; S9, through the incremental learning algorithm, combined with the latest monitoring data and prediction results, the deep learning model parameters are updated in real time; S10. Deploy the optimized deep learning model to the enterprise or power system for real-time carbon emission monitoring and intelligent scheduling.

2. The carbon emission calculation method based on deep learning according to claim 1, characterized in that: In the step S1, data collection includes the fusion of multi-source heterogeneous data, and specifically adopts a generative adversarial network to perform data enhancement. The algorithm formula of the generative adversarial network is as follows: The goal of the generative adversarial network is to enable the generator to generate virtual data similar to the real data distribution through adversarial training between the generator G and the discriminator D. Its loss function is expressed as: Among them, x~p data (x) represents the sample sampled from the real data distribution; z~p z (z) represents random noise sampled from a priori noise distribution; G(z) is a virtual sample generated by the generator; D(x) is the discriminant output probability of the discriminator for the real sample; D(G(z)) is the discriminant output probability of the discriminator for the generated sample; The goal of the generator is to maximize the probability of the discriminator making an error. The loss function of the generator is: The generator optimizes its parameters by minimizing the loss to generate realistic virtual samples; The goal of the discriminator is to maximize the ability to correctly classify real samples and generated samples, and its loss function is: The discriminator optimizes its parameters by minimizing the loss, thereby improving the ability to distinguish between real samples and generated samples.

3. The carbon emission calculation method based on deep learning according to claim 1 is characterized in that: In the S2 step, the autoencoder used for feature extraction is combined with a sparse regularization term during the encoding process so that more useful information is retained in the low-dimensional embedding space. The optimization formula is as follows: Among them, L is the loss function, X i is the input sample, To reconstruct the sample, W j is the model weight, and λ is the regularization coefficient.

4. The carbon emission calculation method based on deep learning according to claim 1 is characterized in that: In the step S3, in the dynamic statistical analysis step, an online Bayesian learning method is used to adaptively adjust the model parameters. The algorithm formula of the online Bayesian learning method is as follows: Bayesian update formula: For parameter θ, the posterior probability update formula is: Among them, P(θ|D t ) is a given data set D t The posterior distribution of P(D t |θ) is the likelihood function; P(θ|D t-1 ) is the posterior distribution of the previous moment; P(D t ) is the marginal likelihood of the data; Definition of likelihood function: For the autoregressive integrated moving average model, assuming that the observed data y t The error is normally distributed, then the likelihood function is defined as: where y i is the actual observed value, is the predicted value of the autoregressive integrated moving average model, 2σ 2 is the variance of the error.

5. The carbon emission calculation method based on deep learning according to claim 1 is characterized in that: In step S5, uncertainty quantification uses a Bayesian neural network to perform probability distribution modeling on the neural network weights to achieve uncertainty assessment of the predicted output. The formula is: P(Y|X)=∫P(Y|W,X)P(W|X)dW, Among them, P(Y|X) represents the output prediction probability, and P(W|X) represents the probability distribution of the weight.

6. The carbon emission calculation method based on deep learning according to claim 1, characterized in that: In step S6, the adaptive adjustment coefficient α of the carbon emission factor i , β i , γ i The objective function proxy model is established through the Bayesian optimization method to find the optimal value of the objective function. The Gaussian process is used to model the objective function, and the model is updated based on the observed data to predict the value of the unobserved point. The algorithm formula is as follows: Gaussian process regression model: Among them, f(x) is the objective function; f(x) is the mean function; k(x,x′) is the kernel function; Probability distribution of the objective function: Given an observed point X = {x1, x2, …, x n } and the corresponding target value y={y1,y2,…,y n }, the new input point x * The target value y * The predicted distribution of is: Among them, K(X,X) is the covariance matrix between observed points, K(x * ,X) is the covariance between the new observation point and the observed point; Update of the posterior distribution: By Bayes' theorem, the new point x * The predicted value y * The posterior mean and variance of μ(x * )=K(x * ,X)K(X,X) -1 y, σ 2 (x * )=K(x * ,x * )-K(x * ,X)K(X,X) -1 K(X,x * ), Among them, μ(x * ) is the predicted mean, σ 2 (x * ) is the prediction variance; Adaptive adjustment coefficient optimization: Calculate and update the adaptive adjustment coefficient α of the carbon emission factor through the Gaussian process model i , β i , γ i ,The goal is to find the optimal value of the coefficient so that the error of predicted carbon emissions is reduced; The Bayesian optimization process determines the next set of test points for the coefficients based on the expected improvement or upper confidence bound strategy, as follows: EI(x) = (μ(x) - y best )Φ(Z)+σ(x)φ(Z), Among them, y best is the optimal value in the observation, Φ(Z) is the cumulative distribution function of the standard normal distribution, and φ(Z) is the probability density function of the standard normal distribution.

7. The carbon emission calculation method based on deep learning according to claim 1 is characterized in that: In the S4 step, the carbon emission prediction model adopts a deep learning model with a hybrid architecture. The convolution layer is used to extract spatial features, the recurrent layer is used to capture time series features, and the Attention mechanism is used to optimize feature weights. The algorithm formula is as follows: Hybrid deep learning model architecture: The model includes a combination of convolutional neural network and recurrent neural network or long short-term memory network, and uses the Attention mechanism to optimize feature weights. The model architecture is as follows: Input layer: Input carbon emission related multidimensional time series data X = {x1, x2, ..., x T }, where each x t is the eigenvector at time t; Convolutional layer: used to extract local spatial features in the data, and perform feature extraction through multiple convolution kernels: Among them, W i is the i-th convolution kernel, * indicates the convolution operation, b i is the bias term; Recurrent layer: used to capture long-term dependencies in time series. The core of the recurrent layer is the cell state and gating mechanism. The update formula of the loop layer is: The information that the forget gate determines to be retained: f t =σ(W f ·[h t-1 ,x t ]+b f ), The input gate decides to update new information: i t =σ(W i ·[h t-1 ,x t ]+b i ), Cell status update: The output gate determines the amount of information output: o t =σ(W o ·[h t-1 ,x t ]+b o ), h t =o t *tanh(C t )。 8. The carbon emission calculation method based on deep learning according to claim 7 is characterized in that: The Attention mechanism is used to assign different importance weights to input features to optimize the use of features. The Attention score calculation formula is: Among them, e t is the attention score; W h , W s is the weight matrix; s t-1 is the hidden state of the previous time step; v is the vector used to generate attention weights; Attention weight calculation: Context vector generation: Context vector c t The hidden states at different time steps are weighted and summed to reflect the importance; Output layer: Use context vector c t Generate the final carbon emissions forecast: Among them, W y is the output layer weight matrix, b y is the output bias.

9. The carbon emission calculation method based on deep learning according to claim 1, characterized in that: In step S5, the hidden Markov model is used to capture the hidden state transition characteristics in the time series. The HMM consists of hidden state, observed state, state transition probability, observation probability and initial probability distribution. The basic elements of the hidden Markov model are: S={s1,s2,…,s N }: hidden state collection; O={o1,o2,…,o T }: observation sequence; A=[a ij ]: state transition probability matrix, a ij =P(s j |s i ); B=[b j (o)]: observation probability matrix, b j (o) = P(o|s j ); π=[π i ]: initial state probability distribution, π i =P(s i ); Forward algorithm: used to calculate the probability P(O|λ) of a given observation sequence; Define the forward variable α t (i) indicates that the state at time t is s i And observe o1,o2,…,o t Probability: α t (i)=P(o1,o2,…,o t ,s i |λ), The recursive formula is: Initialization: α1(i) = π i b i (o1), Recursion: termination:

10. The carbon emission calculation method based on deep learning according to claim 1, characterized in that: In the step S8, the result visualization module can update the key indicators of carbon emission prediction in real time, display uncertainty analysis, trend prediction and sources of abnormal fluctuations, and provide intelligent support for management decisions.

Images