Intelligent prediction method for carbon emission in building operation stage based on neural network optimization

Abstract

The invention relates to a neural network optimization-based intelligent prediction method for carbon emission in a building operation stage. The method comprises the following steps of S1, acquiring multi-dimensional data generated in a building operation process; s2, the obtained multi-dimensional data is preprocessed; s3, constructing a preliminary carbon emission prediction model, and capturing a complex nonlinear relationship in the data by adopting a Swin Transform; s4, optimizing the model by combining an adaptive differential evolution algorithm and a quantum behavior particle swarm optimization algorithm; s5, verifying the optimized model by adopting K-fold cross validation; s6, performing real-time carbon emission prediction by using the optimized model, and outputting a predicted value; and S7, generating a carbon emission analysis report according to a prediction result. According to the method, the enhanced neural network optimization algorithm and real-time dynamic data are combined, high-precision carbon emission prediction and energy consumption management optimization are realized, the prediction accuracy and adaptability are improved, and building energy conservation and emission reduction management is effectively supported.

Description

Technical Field

[0001] The present invention relates to the technical field of building energy conservation and environmental monitoring, and in particular to an intelligent prediction method for carbon emissions in a building operation phase based on neural network optimization. Background Art

[0002] As the global climate change problem becomes increasingly serious, carbon emissions generated during building operation have become one of the important sources of global greenhouse gas emissions. In order to achieve the goal of energy conservation and emission reduction, the monitoring and prediction of building carbon emissions has become an important part of building management. Existing building carbon emission prediction methods are mostly based on linear regression or empirical models. These models rely on static historical data and are difficult to cope with variable factors during building operation, such as real-time changes in environmental conditions, equipment status, and energy usage patterns. In addition, the multidimensionality and complexity of building operation data increase the difficulty of high-precision prediction. Therefore, traditional methods often cannot accurately capture the nonlinear relationship between carbon emissions and these dynamic factors, resulting in insufficient accuracy and timeliness of prediction results.

[0003] Existing technologies have many limitations in processing complex data and dynamic prediction, mainly manifested in: first, traditional prediction models cannot effectively process massive amounts of real-time data, and are prone to ignoring the impact of dynamic data changes on carbon emissions; second, existing models lack intelligent optimization capabilities and cannot dynamically adjust the parameters of the prediction model in a complex building operating environment to adapt to different application scenarios; third, the carbon emission prediction results lack targeted optimization suggestions, making it difficult to provide effective technical support for the actual management of building carbon emissions. Therefore, there is an urgent need for an intelligent prediction method that can make full use of multi-dimensional data and combine neural network optimization algorithms to solve the problems of inaccurate and insufficient real-time prediction of carbon emissions in existing technologies, and thus provide efficient decision-making support for energy conservation and emission reduction in buildings. Summary of the invention

[0004] The technical problem to be solved by the present invention is: in order to overcome the deficiencies in the prior art, a method for intelligently predicting carbon emissions during the operation phase of a building based on neural network optimization is provided.

[0005] The technical solution adopted by the present invention is: a method for intelligent prediction of carbon emissions in the building operation stage based on neural network optimization, comprising the following steps: S1: obtaining multi-dimensional data generated during the building operation process in real time through a sensor network installed in the building, specifically including energy consumption, indoor and outdoor environmental parameters and equipment operation status; the sensor network is equipped with a centralized data aggregation module to improve the real-time and efficiency of data collection and transmission;

[0006] S2: Preprocess the multi-dimensional data obtained by S1. The preprocessing includes filling in missing data, filtering data noise and removing outliers. The method based on long short-term memory network LSTM combined with wavelet transform is used to perform dynamic data filling and outlier detection, and all data are standardized.

[0007] S3: Based on the preprocessed multi-dimensional data and historical carbon emission data, a preliminary carbon emission prediction model is constructed; the preliminary carbon emission prediction model uses the visual architecture Swin Transformer combined with the self-attention mechanism to capture the complex nonlinear characteristics in the building energy consumption data, and the output layer is the predicted carbon emission value;

[0008] Among them, Swin Transformer is a classic visual Transformer architecture designed specifically for computer vision tasks. Swin Transformer was published in ICCV 2021. It is based on the idea of ​​the ViT model and has made many innovations, making it a general backbone network for computer vision and widely used in image classification, object detection, semantic segmentation and other visual tasks.

[0009] S4: Based on the preliminary carbon emission prediction model, the neural network model is optimized by combining the adaptive differential evolution algorithm ADE with the quantum behavior particle swarm optimization algorithm QPSO. The adaptive differential evolution algorithm ADE is used for global parameter search and adaptive adjustment, while the quantum behavior particle swarm optimization algorithm QPSO further performs local search and fine adjustment to improve the optimization effect of the model.

[0010] S5: Validate the optimized neural network model by using the K-fold cross validation method combined with bias-variance analysis to deeply evaluate the generalization ability of the model, divide the preprocessed data set into multiple subsets, and perform model validation step by step;

[0011] S6: Apply the verified neural network model to the real-time carbon emission prediction of buildings and output the predicted building carbon emission value, combined with the threshold control-based strategy to achieve detection and alarm of abnormal carbon emissions;

[0012] S7: Generate a carbon emission analysis report based on the real-time predicted carbon emission value, wherein the analysis report includes carbon emission trends, abnormal emission detection, and carbon emission reduction recommendations based on threshold control, to assist managers in formulating long-term carbon emission reduction strategies.

[0013] Furthermore, the sensor network described in the present invention includes power meters, temperature sensors, humidity sensors, air pressure sensors, carbon dioxide concentration sensors and equipment operation status monitoring sensors. The sensors are distributed in various key locations of the building and combined with a perception coverage optimization mechanism based on heuristic rules to improve the integrity and accuracy of data collection.

[0014] Furthermore, in S2 of the present invention, missing data is supplemented by a long short-term memory network LSTM combined with a dynamic moving average method. By setting an adaptive window size w, the missing value position i in the time series data is interpolated and supplemented using the weighted average of the data of w time points before and after the position i; an anomaly detection method based on cluster analysis is used to eliminate outliers.

[0015] Furthermore, the visual architecture Swin Transformer described in the present invention is a neural network, and its structure is a multi-layer feedforward neural network, including at least three hidden layers. The number of neurons in the hidden layer is adaptively optimized according to the feature dimension and feature importance of the input data to improve the training efficiency and prediction accuracy of the model.

[0016] Furthermore, the specific method of optimizing the neural network model by combining the adaptive differential evolution algorithm ADE with the quantum behavior particle swarm optimization algorithm QPSO in S4 of the present invention is:

[0017] Step A: The adaptive differential evolution algorithm ADE is used for global search of model parameters. By adaptively adjusting the differential variation factor and crossover probability, the adaptability of the model in different feature dimensions is enhanced to ensure that the optimal solution is found in complex building energy consumption data scenarios.

[0018] Step B: The quantum-behavior particle swarm optimization algorithm QPSO is responsible for further local fine-tuning of the solution obtained in step A by introducing quantum behavior mechanisms and using probability distribution for search to better explore the local area of ​​the solution space;

[0019] Step C: Combined with the hybrid adaptive inertia weight strategy, in the quantum-behaved particle swarm optimization, the inertia weight ω decreases nonlinearly with the increase of the iteration number t. The specific formula is as follows:

[0020] Among them, ω max is the initial inertia weight, ω min is the minimum inertia weight, T is the total number of iterations, t is the current number of iterations, and γ is the control factor. This formula uses nonlinear and periodic adjustment strategies to enable particles to dynamically balance global and local search capabilities during the search process, thereby improving the stability and convergence speed of optimization.

[0021] Furthermore, in the K-fold cross validation in S5 of the present invention, the value of K is 10, and the preprocessed data set is evenly divided into 10 subsets, one of the subsets is selected as the validation set each time, and the remaining 9 subsets are used as training sets. The training and validation process is repeated 10 times, and finally the average performance of 10 validations is taken as the generalization ability evaluation index of the model.

[0022] Furthermore, the real-time predicted carbon emission value in S7 of the present invention is used to drive the building energy consumption management system, which dynamically adjusts the equipment operating parameters according to the prediction results. The specific operating parameters include the temperature setting of the HVAC system, the on and off time of the lighting system, and the optimization control of the energy storage system.

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] 1. This invention achieves high-precision intelligent prediction of carbon emissions by introducing an enhanced neural network optimization algorithm and combining multi-dimensional data from the building operation stage. Compared with existing linear or empirical models, this invention can better capture nonlinear relationships in complex dynamic operation data, significantly improving the accuracy and real-time performance of prediction results.

[0025] 2. The present invention uses ADE and QPSO to train the neural network model, so that the model has stronger global search and local optimization capabilities, ensuring the adaptability of the model in different building scenarios, thereby providing more accurate carbon emission prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flowchart of the intelligent prediction method of carbon emissions in the building operation stage in the present invention; DETAILED DESCRIPTION

[0027] The embodiments of the present invention are described in detail below. The embodiments are implemented on the premise of the technical solution of the present invention, and detailed implementation methods and specific operation processes are given, but the protection scope of the present invention is not limited to the following embodiments.

[0028] A method for intelligently predicting carbon emissions during building operation based on neural network optimization includes the following steps (e.g. Figure 1 shown):

[0029] S1: Through the sensor network installed in the building, multi-dimensional data generated during the operation of the building is obtained in real time, including energy consumption, indoor and outdoor environmental parameters (such as temperature, humidity, air pressure, etc.) and equipment operation status; the sensor network is equipped with a centralized data aggregation module to improve the real-time and efficiency of data collection and transmission; the deployment location of the sensors is carefully designed to ensure the comprehensiveness and accuracy of data collection to support subsequent model training and optimization;

[0030] S2: Preprocess the multi-dimensional data obtained by S1. The preprocessing includes completing missing data, filtering data noise and removing outliers. Dynamic data completion and outlier detection are performed based on the long short-term memory network LSTM combined with wavelet transform to improve data quality. All data are standardized so that data from different sources can be analyzed on the same scale, thereby providing consistent data input for subsequent models.

[0031] S3: Based on the preprocessed multi-dimensional data and historical carbon emission data, a preliminary carbon emission prediction model is constructed; the preliminary carbon emission prediction model uses the visual architecture Swin Transformer combined with the self-attention mechanism to capture the complex nonlinear characteristics in the building energy consumption data, and the output layer is the predicted carbon emission value;

[0032] The preliminary carbon emission prediction model uses Swin Transformer, which is a multi-layer feedforward neural network with at least three hidden layers. Swin Transformer uses the window self-attention mechanism to divide the input data into windows of fixed size, calculates self-attention within each window, thereby reducing computational complexity, and combines cross-window connections to capture a wider range of contextual information. The number of neurons in the hidden layer is adaptively optimized according to the feature dimension and feature importance of the input data. The specific optimization process includes using the cross-validation method to select the optimal number of hidden layers and the number of neurons in each layer. The window self-attention mechanism of Swin Transformer enables it to effectively capture the correlation between different dimensions of the input data.

[0033] S4: Based on the preliminary carbon emission prediction model, the neural network model is optimized by combining the adaptive differential evolution algorithm ADE with the quantum behavior particle swarm optimization algorithm QPSO. The adaptive differential evolution algorithm ADE is used for global parameter search and adaptive adjustment, while the quantum behavior particle swarm optimization algorithm QPSO further performs local search and fine adjustment to improve the optimization effect of the model.

[0034] S5: Validate the optimized neural network model by using the K-fold cross validation method combined with bias-variance analysis to deeply evaluate the generalization ability of the model, divide the preprocessed data set into multiple subsets, and perform model validation step by step;

[0035] S6: Apply the verified neural network model to the real-time carbon emission prediction of buildings and output the predicted building carbon emission value, combined with the threshold control-based strategy to achieve detection and alarm of abnormal carbon emissions;

[0036] S7: Generate a carbon emission analysis report based on the real-time predicted carbon emission value, wherein the analysis report includes carbon emission trends, abnormal emission detection, and carbon emission reduction recommendations based on threshold control, to assist managers in formulating long-term carbon emission reduction strategies.

[0037] Among them, the sensor network includes power meters, temperature sensors, humidity sensors, air pressure sensors, carbon dioxide concentration sensors and equipment operation status monitoring sensors. The sensors are distributed in various key locations of the building and combined with a perception coverage optimization mechanism based on heuristic rules to improve the integrity and accuracy of data collection.

[0038] In S2, missing data is completed by a long short-term memory network LSTM combined with a dynamic moving average method. By setting an adaptive window size w, the weighted average of the data of w time points before and after position i is used to interpolate and complete the missing value position i in the time series data; an anomaly detection method based on cluster analysis is used to eliminate outliers.

[0039] The visual architecture Swin Transformer is a neural network whose structure is a multi-layer feedforward neural network, including at least three hidden layers. The number of neurons in the hidden layer is adaptively optimized according to the feature dimension and feature importance of the input data to improve the training efficiency and prediction accuracy of the model.

[0040] The specific method of optimizing the neural network model in S4 by combining the adaptive differential evolution algorithm ADE with the quantum behavior particle swarm optimization algorithm QPSO is as follows:

[0041] In the initial stage of global search, ADE randomly generates a population, each individual represents a possible combination of model parameters, usually with a population size of N and a parameter dimension of D. The diversity of the initial population helps cover a larger search space, thus laying the foundation for finding the global optimal solution;

[0042] The mutation operation of ADE enhances adaptability by dynamically adjusting the mutation factor FFF. In traditional differential evolution, the mutation factor F is fixed, while in ADE, the mutation factor is adaptively adjusted according to the fitness, so that the search process can adaptively change the exploration intensity at different iteration stages. The crossover operation generates new candidate solutions by exchanging genes between the mutant individuals and the original individuals using the crossover probability CR. The crossover probability is also adaptively adjusted according to the individual fitness.

[0043] For each candidate solution, ADE decides whether to keep it by comparing its fitness with that of the parent individual. If the new individual has better fitness, it replaces the old individual, thereby gradually improving the quality of the entire population.

[0044] QPSO performs local search and fine adjustment based on the optimal solution found by ADE. The specific steps are as follows:

[0045] In QPSO, the state of a particle is described by a probability wave function, and the specific position of the particle is obtained through ADE;

[0046] The adaptive inertia weight strategy is adopted in QPSO, and the inertia weight w decreases nonlinearly with the increase of iteration number t to further improve the convergence performance of the algorithm. The specific expression is:

[0047] Among them, ω max is the initial inertia weight, ω min is the minimum inertia weight, T is the total number of iterations, t is the current number of iterations, and γ is the control factor.

[0048] In the K-fold cross-validation in S5, the value of K is 10. Specifically, the preprocessed dataset is evenly divided into 10 subsets, one of the subsets is selected as the validation set each time, and the remaining 9 subsets are used as training sets. The training and validation process is repeated 10 times, and the average performance of the 10 validations is finally taken as the generalization ability evaluation indicator of the model.

[0049] The real-time predicted carbon emission values ​​in S7 are used to drive the building energy management system, which dynamically adjusts the equipment operating parameters based on the predicted results. The specific operating parameters include the temperature setting of the HVAC system, the on and off time of the lighting system, and the optimization control of the energy storage system.

[0050] The carbon emission analysis report includes carbon emission trend analysis, abnormal emission detection and early warning information. The specific contents include:

[0051] Carbon emission trend analysis is used to draw a time series graph of carbon emissions based on historical data and forecast results, and analyze the growth or decline trend of carbon emissions;

[0052] Abnormal emission detection is used to identify the period or cause of abnormally high carbon emissions by setting thresholds or using statistical methods;

[0053] The system automatically generates warning information when abnormal emissions are detected, and provides corresponding processing suggestions, such as equipment maintenance or adjustment of operating strategies.

Claims

1. A method for intelligent prediction of carbon emissions in the building operation phase based on neural network optimization, characterized by: The following steps are involved: S1: Through the sensor network installed in the building, multi-dimensional data generated during the operation of the building is obtained in real time, including energy consumption, indoor and outdoor environmental parameters and equipment operation status; the sensor network is equipped with a centralized data aggregation module to improve the real-time and efficiency of data collection and transmission; S2: Preprocess the multi-dimensional data obtained by S1. The preprocessing includes filling in missing data, filtering data noise and removing outliers. The method based on long short-term memory network LSTM combined with wavelet transform is used to perform dynamic data filling and outlier detection, and all data are standardized. S3: Based on the preprocessed multi-dimensional data and historical carbon emission data, a preliminary carbon emission prediction model is constructed; the preliminary carbon emission prediction model uses the visual architecture Swin Transformer combined with the self-attention mechanism to capture the complex nonlinear characteristics in the building energy consumption data, and the output layer is the predicted carbon emission value; S4: Based on the preliminary carbon emission prediction model, the neural network model is optimized by combining the adaptive differential evolution algorithm ADE with the quantum-behaved particle swarm optimization algorithm QPSO. The adaptive differential evolution algorithm ADE is used for global parameter search and adaptive adjustment, while the quantum behavior particle swarm optimization algorithm QPSO further performs local search and fine adjustment to improve the optimization effect of the model; S5: Validate the optimized neural network model by using the K-fold cross validation method combined with bias-variance analysis to deeply evaluate the generalization ability of the model, divide the preprocessed data set into multiple subsets, and perform model validation step by step; S6: Apply the verified neural network model to the real-time carbon emission prediction of buildings and output the predicted building carbon emission value, combined with the threshold control-based strategy to achieve detection and alarm of abnormal carbon emissions; S7: Generate a carbon emission analysis report based on the real-time predicted carbon emission value, wherein the analysis report includes carbon emission trends, abnormal emission detection, and carbon emission reduction recommendations based on threshold control, to assist managers in formulating long-term carbon emission reduction strategies.

2. The method for intelligent prediction of carbon emissions in the building operation phase based on neural network optimization according to claim 1 is characterized by: The sensor network includes power metering meters, temperature sensors, humidity sensors, air pressure sensors, carbon dioxide concentration sensors and equipment operation status monitoring sensors. The sensors are distributed in various key locations of the building and combined with a perception coverage optimization mechanism based on heuristic rules to improve the integrity and accuracy of data collection.

3. The method for intelligent prediction of carbon emissions in the building operation phase based on neural network optimization according to claim 1 is characterized by: In S2, missing data is completed by a long short-term memory network LSTM combined with a dynamic moving average method. By setting an adaptive window size w, the weighted average of the data of w time points before and after position i is used to interpolate and complete the missing value position i in the time series data; an anomaly detection method based on cluster analysis is used to eliminate outliers.

4. The method for intelligent prediction of carbon emissions in the building operation phase based on neural network optimization according to claim 1 is characterized by: The visual architecture Swin Transformer is a neural network whose structure is a multi-layer feedforward neural network, including at least three hidden layers. The number of neurons in the hidden layer is adaptively optimized according to the feature dimension and feature importance of the input data to improve the training efficiency and prediction accuracy of the model.

5. The method for intelligent prediction of carbon emissions in the building operation phase based on neural network optimization according to claim 1 is characterized by: The specific method of optimizing the neural network model in S4 by combining the adaptive differential evolution algorithm ADE with the quantum behavior particle swarm optimization algorithm QPSO is as follows: Step A: The adaptive differential evolution algorithm ADE is used for global search of model parameters. By adaptively adjusting the differential variation factor and crossover probability, the adaptability of the model in different feature dimensions is enhanced to ensure that the optimal solution is found in complex building energy consumption data scenarios. Step B: The quantum-behavior particle swarm optimization algorithm QPSO is responsible for further local fine-tuning of the solution obtained in step A by introducing quantum behavior mechanisms and using probability distribution for search to better explore the local area of ​​the solution space; Step C: Combined with the hybrid adaptive inertia weight strategy, in the quantum-behaved particle swarm optimization, the inertia weight ω decreases nonlinearly with the increase of the iteration number t. The specific formula is as follows: Among them, ω max is the initial inertia weight, ω min is the minimum inertia weight, T is the total number of iterations, t is the current number of iterations, and γ is the control factor. This formula uses nonlinear and periodic adjustment strategies to enable particles to dynamically balance global and local search capabilities during the search process, thereby improving the stability and convergence speed of optimization.

6. The method for intelligent prediction of carbon emissions during building operation based on neural network optimization according to claim 1 is characterized by: In the K-fold cross-validation in S5, the value of K is 10. Specifically, the preprocessed dataset is evenly divided into 10 subsets, one of the subsets is selected as the validation set each time, and the remaining 9 subsets are used as training sets. The training and validation process is repeated 10 times, and the average performance of the 10 validations is finally taken as the generalization ability evaluation indicator of the model.

7. The method for intelligent prediction of carbon emissions in the building operation phase based on neural network optimization according to claim 1 is characterized by: The real-time predicted carbon emission values ​​in S7 are used to drive the building energy management system, which dynamically adjusts the equipment operating parameters based on the predicted results. The specific operating parameters include the temperature setting of the HVAC system, the on and off time of the lighting system, and the optimization control of the energy storage system.

Images