Cold region building operation carbon emission prediction method based on neural network

Abstract

The invention discloses a cold region building operation carbon emission prediction method based on a neural network, and belongs to the field of cold region building operation carbon emission prediction. The problem of poor accuracy of existing cold region building carbon emission prediction is solved. The method comprises the following steps: collecting power consumption and heat consumption of a current to-be-tested building, calculating operation carbon emission data of the current to-be-tested building by adopting a carbon emission coefficient method, inputting the operation carbon emission data into a pre-trained long-short term neural network model, and obtaining a predicted value of the carbon emission of the to-be-tested building. The method is mainly used for predicting the operation carbon emission of the buildings in the cold regions.

Description

Technical Field

[0001] The present invention belongs to the field of carbon emission prediction for cold-region building operation. Background Art

[0002] The challenges brought about by global warming are becoming increasingly severe, and carbon emission control has become one of the primary tasks of governments and research institutions in various countries. The cold region refers to the severe cold region and the cold region in the thermal zoning. Due to the large heating demand, long-term heating is an important source of carbon emissions in buildings in cold climate regions. Therefore, improving the accuracy of its carbon emission prediction helps to formulate effective emission reduction measures and achieve the carbon neutrality goal. Currently, there are various existing building carbon emission prediction methods based on neural networks, but there are still significant deficiencies in predicting the carbon emissions of cold-region building operation, especially in not considering the complexity of the carbon time-series data of cold-region building operation, including: the significant seasonal differences in the energy consumption patterns during the building operation stage, and the existing methods do not distinguish and consider the energy use differences between the heating period and the non-heating period. Therefore, there is a problem of poor accuracy in carbon emission prediction. Summary of the Invention

[0003] The present invention is to solve the problem of poor accuracy in the existing carbon emission prediction of cold-region buildings, and proposes a method for predicting the carbon emissions of cold-region building operation based on a neural network.

[0004] The method for predicting the carbon emissions of cold-region building operation based on a neural network according to the present invention includes: collecting the power consumption and heat consumption of the currently to-be-tested building, and using the carbon emission coefficient method to calculate the carbon emissions generated by the power consumption during the current building operation and the carbon emissions generated by the heat consumption of the heating system respectively, and inputting the carbon emission data generated by the power consumption and the carbon emission data generated by the heat consumption into a pre-trained long short-term neural network model to obtain the predicted value of the carbon emissions of the to-be-tested building operation.

[0005] The long short-term neural network model is trained using the sample data of the historical power consumption and heat consumption of the to-be-tested building. The specific training method is as follows:

[0006] Step 1: Using the carbon emission coefficient method, and using the pre-processed power consumption and heat consumption data, calculate the carbon emission data generated by the power consumption of the to-be-tested building and the carbon emission data generated by the heating heat consumption respectively;

[0007] Step 2: Using the Pearson product-moment correlation coefficient, perform a correlation analysis on the carbon emissions generated by the power consumption and meteorological factors to obtain the correlation coefficient;

[0008] Step 3: Divide the carbon emission data generated by the power consumption into a strong correlation group, a medium correlation group, and a weak correlation group according to the correlation coefficient. Then, use the outdoor weather comfort index to re-divide the electricity meter data in the medium correlation group to obtain a comfort group and a discomfort group;

[0009] Step 4: Establish four long short-term neural network models. Input the carbon emission data generated by the power consumption in the strong correlation group, weak correlation group, comfort group, and discomfort group into the four long short-term neural network models respectively to predict the carbon emissions generated by the power consumption, and obtain the prediction results of the carbon emissions generated by the power consumption;

[0010] Meanwhile, establish a fifth long short-term neural network model. Input the carbon emission data generated by the heat consumption into the fifth long short-term neural network model to predict the carbon emissions generated by the heating heat consumption, and obtain the carbon emission results generated by the heating heat consumption;

[0011] Sum up the prediction results of the carbon emissions generated by the power consumption and the prediction results of the carbon emissions generated by the heating heat consumption to obtain the final prediction result of the building operation carbon emissions.

[0012] Furthermore, in the present invention, in Step 2, the method for calculating the carbon emission data generated by the power consumption of the building to be tested and the carbon emission data generated by the heating heat consumption by using the carbon emission coefficient method and the preprocessed power consumption and heat consumption data is as follows:

[0013] Use the formula:

[0014] CE = BFS × CEI

[0015] Calculate the building operation carbon emissions CE, where BFS is the building area of the test building and CEI is the building operation carbon emission intensity;

[0016] The building operation carbon emission intensity is:

[0017]

[0018] Among them, EUI is the energy use index of each energy, CEF represents the carbon emission factor of each energy, and S represents the number of energies participating in the operation carbon emissions;

[0019] The calculation of the carbon emission data generated by the power consumption is:

[0020] CEI 电 = EUI 电 × CEF 电

[0021] EUI 电 is the energy use index of the electric energy, CEF 电is the carbon emission factor of electric energy;

[0022] The calculated carbon emission data generated by heating heat energy consumption is:

[0023] CEI 热 = EUI 热 × CEF 热 .

[0024] EUI 热 is the energy use index of heating heat energy, and CEF 热 is the carbon emission factor of heating heat energy.

[0025] Furthermore, in the present invention, in step three, the method for performing a correlation analysis on the carbon emission data generated by electric energy consumption and meteorological factors to obtain the correlation coefficient is as follows:

[0026] Divide the carbon emission data generated by the annual electric energy consumption into 12 time periods with a month as the basic unit, and perform a correlation analysis on the carbon emission data generated by the building's electric energy consumption and meteorological factors in each time period;

[0027] Pearson product-moment correlation coefficient PCCs:

[0028]

[0029] where cov(X,Y) is the covariance of the rank variables, and σ X σ Y is the standard deviation of the rank variables, and X and Y are the carbon emission data generated by the building's electric energy consumption and the average temperature respectively.

[0030] Furthermore, in the present invention, in step four, the method for dividing the carbon emission data generated by the electric energy consumption into a strong correlation group, a medium correlation group, and a weak correlation group according to the correlation coefficient is as follows:

[0031] According to the absolute value of the Pearson product-moment correlation coefficient PCCs, the correlation is divided into four levels: a strong correlation group, a medium correlation group, and a weak correlation group. The weak correlation group includes weak correlation and non-correlation data;

[0032] Strong correlation group. The absolute value range of the correlation coefficient of the strong correlation group is 0.5 to 1.0, the absolute value range of the correlation coefficient of the medium correlation group is 0.3 to 0.5, and the absolute value range of the correlation coefficient of the weak correlation group is 0 to 0.3.

[0033] Furthermore, in the present invention, in step four, the method for obtaining a comfortable group and an uncomfortable group is as follows:

[0034] Re-group the medium correlation group using the outdoor weather comfort index OWCI;

[0035] OWCI = C - (C ta + C tm + C h + C v + C r )

[0036] where C is a constant, C ta is the average temperature term, C tm is the maximum temperature term, C h is the humidity term, C v is the wind speed term, C r is the radiation term. The data with a comfort index ≥ 7.0 are divided into the comfortable group, and the data with a comfort index ≤ 7.0 are divided into the uncomfortable group.

[0037] Furthermore, in the present invention, in step five, the long short-term neural network model includes an output gate (1), a forget gate (2), an input gate (3), a matrix calculation module (4), a first multiplier (5), a first adder (6), a second multiplier (7), and a third multiplier (8);

[0038] The input data is simultaneously output to the output gate (1), the forget gate (2), the input gate (3), and the matrix calculation module (4);

[0039] The input gate (3) controls the input switch using an activation function to control whether the input data is input to the third multiplier (8);

[0040] The forget gate (2) controls whether the predicted value data of the previous time step is retained using an activation function; the retained data is input to the second multiplier (7);

[0041] The output gate (1) is used to control whether the input data of the current time step and the predicted value of the previous time step are converted into a hidden state, and the data that is not converted into a hidden state is transmitted to the first multiplier (5);

[0042] The matrix calculation module (4) is used to perform matrix calculation on the input data to obtain a calculation result, and the calculation result is input to the third multiplier (8);

[0043] The formula for the matrix calculation is:

[0044]

[0045] where C t-1 represents the cell state of the previous time step, f t represents the output of the forget gate, i t represents the input of the input gate, represents the cell candidate value;

[0046]

[0047] Among them, tanh is the activation function, which restricts the range of candidate values to be between [-1, 1], and h t-1 represents the hidden state data of the previous time step, and x t represents the input data, where W c represents the weight matrix of the input unit x t , and U c is the weight matrix of the linear transformation of the hidden state h t-1 of the previous time step, and b c is the bias value, which ensures that the output can be activated even when the input and the hidden state are zero;

[0048] The third multiplier (8) performs element-wise matrix multiplication on the data output by the input gate (3) and the calculation result of the matrix calculation module (4) to obtain a set of candidate cell state values; the set of candidate cell state values is transmitted to the first adder (6);

[0049] The second multiplier (7) performs element-wise matrix multiplication on the result output by the first adder (6) at the previous time step and the data retained by the forget gate (2) to obtain a product result; the product result is input to the first adder (6), and the first adder (6) performs element-wise matrix addition on the product result and the set of candidate cell state values, and inputs the result of the element-wise addition to the second multiplier (7) and the first multiplier (5);

[0050] The first multiplier (5) performs element-wise matrix multiplication on the data that the output gate (1) has not converted into the hidden state and the result of the element-wise addition to obtain the cell state at the next time step; the cell state at the next time step is output as the predicted value.

[0051] Furthermore, in the present invention, the formula for the forget gate (2) to use the activation function to control whether the predicted value data of the previous time step is retained is:

[0052] f t =σ(W f x t +U f h t-1 +b f )

[0053] f t represents the output of the forget gate, σ represents the activation function, and the value is 0 or 1. 0 represents complete forgetting, and 1 represents complete retention; U f represents the weight matrix of the linear transformation of the forget gate for the hidden state h t-1 of the previous time step, W f represents the input weight matrix of the forget gate, and b f represents the bias of the forget gate.

[0054] Furthermore, in the present invention, the input gate (3) controls the input switch by using an activation function, and the formula for controlling the input data is as follows:

[0055] i t =σ(W i x t +U i h t-1 +b i )

[0056] where σ is the activation function, with a value of 0 or 1, where 0 represents closed and 1 represents fully open; it is used to control the switch of the input gate, W i represents the input weight matrix of the input gate, U i represents the linear transformation weight matrix of the input gate for the hidden state h t-1 at the previous time step, b i represents the bias of the input gate, and i t represents the output of the input gate.

[0057] Furthermore, in the present invention, the formula for the output gate (1) to perform a partial state transformation on the input data at the current time step and the predicted value at the previous time step is as follows:

[0058] o t =σ(W o x t +U o h t-1 +b o )

[0059] o t represents the output of the output gate, W o represents the input weight matrix of the output gate, U o represents the linear transformation weight matrix of the output gate for the hidden state h t-1 at the previous time step, and b o represents the bias of the output gate.

[0060] Furthermore, in the present invention, the predicted value is:

[0061] h t =o t ·tanh(C t )

[0062] where C t represents the cell state at time step t.

[0063] In view of the differences in energy use between the heating period and the non - heating period of office buildings in cold regions, the present invention proposes a carbon emission prediction model based on the LSTM neural network to make up for the deficiency in the prior art of insufficient consideration of the seasonal energy use pattern in cold regions. In terms of carbon emission prediction, the carbon emissions of building operation are simulated in groups to reduce the errors caused by data fluctuations. The improved model can effectively cope with the complex energy use characteristics in cold climate regions, especially in scenarios with significant seasonal fluctuations, improve the accuracy of carbon emission prediction for the operation of cold - region buildings, and contribute to promoting the achievement of the regional carbon neutrality goal. Description of the Drawings

[0064] Figure 1 It is a flow chart of the training method of the long - short - term neural network model described in the present invention;

[0065] Figure 2 It is a schematic diagram of the long - short - term neural network model. Detailed Embodiments

[0066] Detailed Embodiment 1: In combination with Figure 1 and Figure 2 This embodiment is described. The method for predicting the carbon emissions of the operation of cold - region buildings based on a neural network includes: collecting the power consumption and heat consumption of the currently tested building, and using the carbon emission coefficient method to calculate the carbon emissions generated by the electricity consumption and the carbon emissions generated by the heat consumption of the heating system in the current operation of the building respectively. Input the carbon emission data generated by the electricity consumption and the carbon emission data generated by the heat consumption into the pre - trained long - short - term neural network model to obtain the predicted value of the carbon emissions of the tested building operation.

[0067] The long - short - term neural network model is trained using the historical power consumption and heat consumption sample data of the tested building. The specific training method is as follows:

[0068] Step 1: Using the carbon emission coefficient method, and using the power consumption and heat consumption sample data, calculate the carbon emission data generated by the electricity consumption and the carbon emission data generated by the heating heat consumption of the tested building respectively;

[0069] Step 2: Using the Pearson product - moment correlation coefficient, conduct a correlation analysis on the carbon emissions generated by the electricity consumption and meteorological factors to obtain the correlation coefficient;

[0070] Step 3: Divide the carbon emission data generated by the electricity consumption into a strong - correlation group, a medium - correlation group, and a weak - correlation group according to the correlation coefficient, and use the outdoor weather comfort index to re - divide the electricity meter data in the medium - correlation group to obtain a comfort group and an uncomfortable group;

[0071] Step 4: Establish four long short-term neural network models. Input the carbon emission data generated from the electricity consumption of the strong correlation group, weak correlation group, comfort group, and discomfort group into the four long short-term neural network models respectively to predict the carbon emissions generated from electricity consumption, and obtain the prediction results of the carbon emissions generated from electricity consumption.

[0072] Meanwhile, establish a fifth long short-term neural network model. Input the carbon emission data generated from heat consumption into the fifth long short-term neural network model to predict the carbon emissions generated from heating heat consumption, and obtain the prediction results of the carbon emissions generated from heating heat consumption.

[0073] Sum up the prediction results of the carbon emissions generated from electricity consumption and the prediction results of the carbon emissions generated from heating heat consumption to obtain the final prediction results of the building operation carbon emissions.

[0074] The historical power consumption and heat consumption sample data of the building to be tested in the present invention are the historical power consumption and heat consumption of the building to be tested at least one hour per day for at least one year, and the power consumption and heat consumption have been preprocessed.

[0075] The data collection interval of the electricity meter data and heat meter data adopted in the present invention should be one hour, and the data time span should be at least one year. Summarize the electricity meter and heat meter data, and conduct itemized sorting in combination with the time series. Missing data can be filled using statistical methods such as linear interpolation method, but the continuous time of missing data should not exceed 24 hours.

[0076] The building operation carbon emission data described in the present invention include: the building carbon emission data caused by electricity consumption and the building carbon emission data caused by heat consumption; during the non-central heating period, the daily building operation carbon is the prediction result of the carbon emissions generated from electricity consumption, and the building carbon emission data caused by heat consumption at this time is 0; during the central heating period, the daily building operation carbon is the sum of the prediction results of the carbon emissions generated from electricity consumption and the carbon emissions generated from heat consumption.

[0077] Further, in the present invention, in step 2, the method for calculating the carbon emission data generated from the electricity consumption of the building to be tested and the carbon emission data generated from the heating heat consumption of the building to be tested respectively by using the carbon emission coefficient method and the preprocessed power consumption and heat consumption data is as follows:

[0078] Use the formula:

[0079] CE = BFS × CEI

[0080] Calculate the building operation carbon emissions CE, where BFS is the building area of the test building and CEI is the building operation carbon emission intensity.

[0081] The building operation carbon emission intensity is:

[0082]

[0083] Among them, EUI is the energy usage index of each energy source, CEF represents the carbon emission factor of each energy source, and S represents the number of energy sources participating in operating carbon emissions;

[0084] The carbon emission data generated by calculating the electricity consumption is:

[0085] CEI 电 = EUI 电 × CEF 电

[0086] EUI 电 is the energy usage index of the electrical energy, and CEF 电 is the carbon emission factor of the electrical energy;

[0087] The carbon emission data generated by calculating the heating heat energy consumption is:

[0088] CEI 热 = EUI 热 × CEF 热 .

[0089] EUI 热 is the energy usage index of the heating heat energy, and CEF 热 is the carbon emission factor of the heating heat energy.

[0090] Furthermore, in the present invention, in step three, the method for performing a correlation analysis on the carbon emission data generated by the electricity consumption and meteorological factors to obtain the correlation coefficient is:

[0091] The carbon emission data generated by the annual electricity consumption is divided into 12 time periods with a month as the basic unit, and a correlation analysis is performed on the carbon emission data generated by the building electricity consumption and meteorological factors in each time period;

[0092] Pearson product-moment correlation coefficient PCCs:

[0093]

[0094] Among them, cov(X,Y) is the covariance of the rank variables, and σ X σ Y is the standard deviation of the rank variables, and X and Y are respectively the carbon emission data generated by the building electricity consumption and the average temperature.

[0095] Furthermore, in the present invention, in step four, the method for dividing the carbon emission data generated by the electricity consumption into a strong correlation group, a medium correlation group, and a weak correlation group according to the correlation coefficient is:

[0096] According to the absolute value of the Pearson product-moment correlation coefficient PCCs, the correlation is divided into four levels: a strong correlation group, a medium correlation group, and a weak correlation group. The weak correlation group includes weakly correlated and uncorrelated data;

[0097] For the strong correlation group, the absolute value range of the correlation coefficient in the strong correlation group is 0.5 to 1.0, the absolute value range of the correlation coefficient in the medium correlation group is 0.3 to 0.5, and the absolute value range of the correlation coefficient in the weak correlation group is 0 to 0.3.

[0098] Furthermore, in the present invention, in step four, the method for obtaining the comfortable group and the uncomfortable group is as follows:

[0099] The medium correlation group is re-grouped using the outdoor weather comfort index OWCI;

[0100] OWCI = C - (C ta + C tm + C h + C v + C r )

[0101] where C is a constant, C ta is the average temperature term, C tm is the maximum temperature term, C h is the humidity term, C v is the wind speed term, C r is the radiation term. Data with a comfort index ≥ 7.0 is divided into the comfortable group, and data with a comfort index ≤ 7.0 is divided into the uncomfortable group.

[0102] Furthermore, in the present invention, in step five, the long short-term neural network model includes an output gate (1), a forgetting gate (2), an input gate (3), a matrix calculation module (4), a first multiplier (5), a first adder (6), a second multiplier (7), and a third multiplier (8);

[0103] The input data is simultaneously output to the output gate (1), the forgetting gate (2), the input gate (3), and the matrix calculation module (4);

[0104] The input gate (3) controls the input switch using an activation function to control whether the input data is input to the third multiplier (8);

[0105] The forgetting gate (2) controls whether the predicted value data of the previous time step is retained using an activation function; the retained data is input to the second multiplier (7);

[0106] The output gate (1) is used to control whether the input data at the current time step and the predicted value of the previous time step are converted into a hidden state, and the data that has not been converted into a hidden state is transmitted to the first multiplier (5);

[0107] The matrix calculation module (4) is used to perform matrix calculations on the input data, obtain the calculation results, and input the calculation results into the third multiplier (8);

[0108] The formula for the matrix calculation is:

[0109]

[0110] where C t-1 represents the cell state at the previous time step, and f t represents the output of the forget gate, and i t represents the input of the input gate, represents the cell candidate value;

[0111]

[0112] where, tanh is the activation function, which limits the range of the candidate value between [-1, 1], and h t-1 represents the hidden state data at the previous time step, and x t represents the input data, where W c represents the weight matrix of the input unit x t and U c is the weight matrix of the linear transformation of the hidden state h t-1 at the previous time step, and b c is the bias value, which ensures that the output can be activated even when the input and the hidden state are zero;

[0113] The third multiplier (8) performs element-wise matrix multiplication on the data output by the input gate (3) and the calculation results of the matrix calculation module (4) to obtain a set of cell state candidate values; the set of cell state candidate values is transmitted to the first adder (6);

[0114] The second multiplier (7) performs element-wise matrix multiplication on the result output by the first adder (6) at the previous time step and the data retained by the forget gate (2) to obtain a product result; the product result is input to the first adder (6), and the first adder (6) performs element-wise matrix addition on the product result and the set of cell state candidate values, and inputs the result of the element-wise addition to the second multiplier (7) and the first multiplier (5);

[0115] The first multiplier (5) performs element-wise matrix multiplication on the data that has not been converted into the hidden state by the output gate (1) and the result of the element-wise addition to obtain the cell state at the next time step; the cell state at the next time step is output as the predicted value.

[0116] Further, in the present invention, the formula for the forget gate (2) to use the activation function to control whether to retain the predicted value data at the previous time step is:

[0117] f t = σ(W f x t + U f h t-1 + b f )

[0118] f t represents the output of the forget gate, σ represents the activation function, with a value of 0 or 1, where 0 represents complete forgetting and 1 represents complete retention; U f represents the linear transformation weight matrix of the forget gate for the hidden state h t-1 at the previous time step, W f represents the input weight matrix of the forget gate, and b f represents the bias of the forget gate.

[0119] Furthermore, in the present invention, the input gate (3) controls the input switch using an activation function, and the formula for controlling the input data is:

[0120] i t = σ(W i x t + U i h t-1 + b i )

[0121] σ is the activation function, with a value of 0 or 1, where 0 represents closed and 1 represents fully open; it is used to control the switch of the input gate, W i represents the input weight matrix of the input gate, U i represents the linear transformation weight matrix of the input gate for the hidden state h t-1 at the previous time step, b i represents the bias of the input gate, and i t represents the output of the input gate.

[0122] Furthermore, in the present invention, the output gate (1) performs a partial state transformation on the input data at the current time step and the predicted value at the previous time step according to the formula:

[0123] o t = σ(W o x t + U o h t-1 + b o )

[0124] o t represents the output of the output gate, W o represents the input weight matrix of the output gate, U o represents the linear transformation weight matrix of the output gate for the hidden state h t-1 at the previous time step, b oIndicates the deviation of the output gate.

[0125] Further, in the present invention, the predicted value is:

[0126] h t = o t ·tanh(C t )

[0127] Wherein, C t represents the cell state at time step t.

[0128] By relying on meteorological characteristics and outdoor weather numerical indices, the present invention decomposes the complex annual fluctuating building operation carbon into carbon emissions generated by electricity consumption and carbon emissions generated by heat consumption. At the same time, it further decomposes the carbon emission data generated by electricity consumption into four groups of building operation carbon emission data, which are simulated separately and then combined. The prediction accuracy is improved compared with directly simulating the annual data.

[0129] In this embodiment, the long short-term neural network (LSTM) adopted is currently the most commonly used method for predicting building operation carbon emissions. However, in the face of long time series data, it will cause the learning ability of the model to decline and the prediction accuracy to decrease.

[0130] A long time series data often contains multiple cycles. For example, the building operation carbon emission data may fluctuate with fixed cycles in a day, a week, or a year. The differences in electricity consumption between day and night, the working patterns on weekdays and weekends, and the carbon emission peaks in winter and summer are all cycles. If directly training the complete data, the model needs to learn multiple patterns, increasing the training difficulty. In this case, the model will have an incomplete learning of each cycle and will only learn the most significant features while ignoring the details.

[0131] Therefore, after grouping the data, the data complexity can be reduced, the feature correlation can be highlighted, and noise interference can be avoided. After each grouping, the model only needs to process a simple pattern, reducing the complexity and at the same time improving the prediction effect and generalization ability.

[0132] Step S1: Collect the original data.

[0133] The building operation carbon in cold regions mainly includes carbon emissions generated by electricity consumption and carbon emissions generated by heating heat consumption. Therefore, collect the hourly electricity meter data and heat meter data of the test buildings in cold regions, and the data time span is at least one year. Summarize the electricity meter and heat meter data, and conduct itemized sorting in combination with the time series. Missing data can be filled using statistical methods such as linear interpolation, but the continuous time of missing data should not exceed 24 hours.

[0134] Step S2: Calculate the electricity meter data and heat meter data into building operation carbon emission data.

[0135] The carbon emissions generated by electricity consumption and the carbon emissions generated by heat consumption are calculated respectively through the carbon emission factor method.

[0136] The carbon emission factor method mentioned refers to the building operation carbon emission calculation model developed by the Intergovernmental Panel on Climate Change (IPCC). This model combines activity level data and carbon emission factors. The specific building operation carbon emission calculation formula (1) is as follows:

[0137] CE = BFS × CEI

[0138] Where CE is the building operation carbon emissions, BFS is the building area of the test building, and CEI is the building operation carbon emission intensity.

[0139] The calculation formula (2) of the building operation carbon emission intensity is as follows:

[0140]

[0141] Where EUI is the energy use index of each energy, CEF represents the carbon emission factor of each energy, and S represents the number of energies participating in the operation carbon emissions. The building operation carbon emission intensity is obtained by summing the products of the energy use index and the carbon emission factor of each energy.

[0142] When calculating the carbon emission data caused by electricity consumption, the formula is as follows:

[0143] CEI 电 = CUI 电 × CEF 电 .

[0144] When calculating the carbon emission data caused by heat consumption, the formula is as follows:

[0145] CEI 热 = CUI 热 × CEF 热 .

[0146] The electricity carbon emission factor (CEF) adopts the average carbon dioxide emission factor of electricity in the Northeast region announced by the National Bureau of Statistics in 2021, and the value is 0.6012 (kgCO2 / kWh).

[0147] The specific building operation carbon emission calculation model is shown in the following figure:

[0148] Step S3: Preliminary grouping, and correlation analysis is carried out through the Pearson product-moment correlation coefficient.

[0149] Next, the operations from step three to step six are only implemented for the carbon emissions generated by electricity consumption, and no operations are performed on the carbon emissions generated by heat consumption.

[0150] Using Pearson product - moment correlation coefficients (PCCs), perform a correlation analysis on the carbon emission data generated by the building's electricity consumption to be tested and meteorological factors.

[0151] The calculation formula for the Pearson product - moment correlation coefficient (PCCs) is as follows:

[0152]

[0153] where cov(X,Y) is the covariance of the ranked variables, and σ X σ Y is the standard deviation of the ranked variables.

[0154] Divide the carbon emissions generated by the building's annual electricity consumption into 12 time periods with a month as the basic unit, and perform a correlation analysis on the carbon emissions generated by the building's electricity consumption and meteorological factors in each time period.

[0155] According to the magnitude of the absolute value of the Pearson product - moment correlation coefficient (PCCs), the correlation is divided into four levels: strong, medium, weak, and none. The classification criteria are shown in Table 1:

[0156] Table 1. Pearson correlation coefficient classification criteria

[0157]

[0158] Step S4: (Group according to the index, extract the data groups with obvious eigenvalues according to the correlation of meteorological factors, and re - group the data with a "medium" correlation value, that is, data with some eigenvalues but not particularly strong.)

[0159] Through the correlation calculation in step S3, divide the carbon emission data generated by the building's annual electricity consumption into three groups G1, G2, and T. G1 is the time period when the carbon emission data generated by the building's electricity consumption has a weak or no correlation with meteorological factors, that is, the time period when the range of |PCC S | is in the range of 0.0 - 0.3; G2 is the time period when the carbon emission data generated by the building's electricity consumption has a strong correlation with meteorological factors, that is, the time period when the range of |PCC S | is in the range of 0.5 - 1.0; T is the time period when the carbon emission data generated by the building's electricity consumption has a medium correlation with meteorological factors, that is, the time period when the range of |PCC S | is in the range of 0.3 - 0.5.

[0160] Among them, for the naming of the groups, G represents Group, and T represents Transition.

[0161] Step S5: For the T grouping in the correlation between the carbon emissions generated by building electricity consumption and meteorological factors in step S4, the Outdoor Weather Comfort Index (OWCI) is used for refined grouping. The calculation formula of the Outdoor Weather Comfort Index (OWCI) is as follows:

[0162] OWCI = C - (C ta + C tm + C h + C v + C r )

[0163] Among them, C is a constant, C ta is the average temperature term, C tm is the maximum temperature term, C b is the humidity term, C v is the wind speed term, C r is the radiation term.

[0164] For the part of the correlation between the carbon emissions data generated by building electricity consumption and meteorological data, this data time often appears in the transitional months. In the cold region, the temperature fluctuates greatly during the transitional season. Therefore, OWCI, that is, the Outdoor Weather Comfort Index, is used for analysis, and features are extracted for more detailed grouping.

[0165] The existing Outdoor Weather Comfort Index (OWCI) can correspond to different comfort index classifications. There are four comfort index classifications in total, namely: extremely comfortable, comfortable, uncomfortable, and extremely uncomfortable. The specific temperature indicators and index values for each level are specified in Table 2 as follows:

[0166] Table 2 OWCI Classification Temperature Indicators and Their Index Values

[0167]

[0168] Among them, T a is the daily average temperature, T max is the daily maximum temperature.

[0169] Calculate the Outdoor Weather Comfort Index daily for the time period T in the correlation between the carbon emissions generated by building electricity consumption and meteorological factors in step four, and form two building operation carbon emission time periods, T1 and T2, according to the results.

[0170] Among them, T1 refers to the time period with the output result of code 1, and T2 refers to the time period with the output result of code 3. If the output code is 2 or 4, then this day has the same grouping time period as the previous day.

[0171] Step S6: According to steps S3, S4, and S5, the carbon emissions generated by the annual building electricity consumption can be divided into four time periods, G1, G2, T1, and T 2；

[0172] After steps S3, S4, and S5, four groups of carbon emission data generated from building electricity consumption can be obtained. These four groups of data cover the whole year's data and have obvious periodicity and characteristic values.

[0173] Through the above steps, based on meteorological factors and outdoor weather comfort index, the carbon emission data generated from building electricity consumption throughout the year is decomposed from a group of complex data into four groups of simple data, and long short-term neural network models (LSTM) are established for the four groups of simple data respectively. At the same time, a long short-term neural network (LSTM) is established separately for the carbon emissions generated from heat energy consumption, and the model training is carried out separately for the carbon emissions generated from heat energy consumption, so as to propose a new time series prediction framework, which is composed of the above 5 data models.

[0174] The long short-term neural network model (LSTM) is a variant of the traditional recurrent neural network (RNN). The long short-term neural network (LSTM) adds a gating system in the memory unit, namely: forget gate, input gate, and output gate. The specific memory unit structure is as follows Figure 2 shown;

[0175] In time series prediction, "time step" usually refers to the time interval between one observation point and the next in the sequence data. It defines the sampling frequency and time scale of the time series data. The operation on the current time step is indicated by the subscript t, and the content of the previous time step is indicated by the subscript t-1.

[0176] The input data in the figure refers to the data x of the current time step currently input to the LSTM unit t ;

[0177] The role of the forget gate is to determine whether historical information is retained, and its formula is:

[0178] f t =σ(W f ·x t +U f .h t-1 +b f )

[0179] In the formula, σ is the activation function, and the value is generally 0 or 1. 0 represents complete forgetting, and 1 represents complete retention.

[0180] The role of the input gate is to determine which parts of the input data at the current time step are written, and its formula is:

[0181] i t =σ(W i ·x t +U i·h t-1 +b i )

[0182] In the formula, σ is the activation function, which is used to control the opening and closing degree of the input gate.

[0183] The function of the output gate is to determine which parts in the current time step are output to the hidden state, and its formula is:

[0184] o t =σ(W o ·x t +U o ·h t-1 +b o )

[0185] In the formula, σ is the activation function σ, which is used to control the opening and closing of the output gate.

[0186] In the above formula, W f 、W i 、W o are the weight matrices of the input unit x t of each gate respectively. This kind of transformation can enable the long short-term neural network to extract complex patterns from the input features; U f 、U i 、U o are the weight matrices of the linear transformation of the hidden state h t-1 of the previous time step. This kind of transformation allows the long short-term neural network to consider historical information when determining the current state;

[0187] h t-1 is the hidden state, that is, the output of each previous time step of the long short-term neural network, and at the same time serves as one of the inputs of the current time step. Its function is to serve as a kind of dynamic memory to transfer short-term information between time steps. b f 、b i and b o represent the biases of the gates, and their function is to help adjust the activation values of the gates, so that the output of the gates can be activated even when the input and the hidden state are zero.

[0188] In the figure and the multiplication sign inside represents element-wise matrix multiplication, and the plus sign inside represents element-wise matrix addition. Among them, the update formula of the cell state is:

[0189]

[0190] Among them, f t ·C t-1 is to multiply the cell state C t-1 of the previous time step by the output of the forgetting gate f t is the product of the input gate i t and the cell candidate value . The cell state candidate value refers to the input data x t and the hidden state h t-1 at the previous time step. Starting from these, the result is generated through a set of weight calculations, and its generation formula is:

[0191]

[0192] where tanh is the activation function, which limits the range of the candidate value between [-1, 1] to prevent the value from being too large or too small.

[0193] Finally, the predicted value in the figure is the final output h t at the current time step, and it also serves as the hidden state for the next time step. Its formula is:

[0194] h t = o t ·tanh(C t )

[0195] The above-mentioned gating system can control the flow of information in the neural network, thereby realizing the processing of short-term and long-term dependencies and effectively solving time series-based problems. Therefore, a long short-term neural network (LSTM) is used for simulation and prediction.

[0196] Step S7: Perform the training of the neural network. Relevant parameters for the training of the long short-term neural network model are set in advance.

[0197] The content of the input value in the long short-term neural network is a matrix with three dimensions (m, T x , n x ).

[0198] Among them, m is the number of data samples in each group, T x is the number of time steps in each sample, and n x is the number of input features.

[0199] The grouped data is divided into a training group, a validation group, and a test group in a ratio of 7:2:1, and the long short-term neural network (LSTM) is trained. The training is stopped before overfitting to obtain the trained neural network model.

[0200] The specific parameter settings of the model are as follows:

[0201] Table 3 Parameter values for the long short-term neural network model settings

[0202]

[0203] Step S8: Calculate the building operation carbon; the carbon emissions generated by electricity consumption and the carbon emissions generated by heating heat consumption are predicted by the trained model, and the predicted results overlap in some time periods. The specific overlapping time is the central heating period of each year. Thus, the building operation carbon prediction result can be obtained as follows:

[0204] During the non - central heating period, the building operation carbon is the predicted result of the carbon emissions generated by electricity consumption;

[0205] During the central heating period, the building operation carbon is the sum of the predicted results of the carbon emissions generated by electricity consumption and the carbon emissions generated by heating heat consumption.

[0206] Although the present invention has been described with reference to specific embodiments herein, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed, as long as they do not depart from the spirit and scope of the present invention as defined by the appended claims. It should be understood that different dependent claims and the features described herein can be combined in a manner different from that described in the original claims. It should also be understood that the features described in connection with a single embodiment can be used in other described embodiments.

Claims

1. A method for predicting the operating carbon emissions of cold-region buildings based on a neural network, characterized in that, Including: Collect the power consumption and heat consumption of the building to be tested currently. Using the carbon emission coefficient method, calculate the carbon emissions generated by the electricity consumption and the carbon emissions generated by the heat consumption of the heating system during the current operation of the building respectively. Input the carbon emission data generated by the electricity consumption and the carbon emission data generated by the heat consumption into a pre-trained long short-term neural network model to obtain the predicted value of the operating carbon emissions of the building to be tested. The long short-term neural network model is trained using the historical power consumption and heat consumption sample data of the building to be tested. The specific training method is as follows: Step 1: Using the carbon emission coefficient method, and using the preprocessed power consumption and heat consumption data, calculate the carbon emission data generated by the electricity consumption of the building to be tested and the carbon emission data generated by the heating heat consumption respectively; Step 2: Use the Pearson product-moment correlation coefficient to perform a correlation analysis on the carbon emissions generated by the electricity consumption and meteorological factors to obtain the correlation coefficient; Step 3: Divide the carbon emission data generated by the electricity consumption into a strong correlation group, a medium correlation group, and a weak correlation group according to the correlation coefficient. Use the outdoor weather comfort index to re-divide the electricity meter data in the medium correlation group to obtain a comfort group and an uncomfortable group; Step 4: Establish four long short-term neural network models. Input the carbon emission data generated by the electricity consumption of the strong correlation group, the weak correlation group, the comfort group, and the uncomfortable group into the four long short-term neural network models respectively to predict the carbon emissions generated by the electricity consumption, and obtain the predicted result of the carbon emissions generated by the electricity consumption; At the same time, establish a fifth long short-term neural network model. Input the carbon emission data generated by the heat consumption into the fifth long short-term neural network model to predict the carbon emissions generated by the heating heat consumption, and obtain the carbon emission result generated by the heating heat consumption; Sum the predicted result of the carbon emissions generated by the electricity consumption and the predicted result of the carbon emissions generated by the heating heat consumption to obtain the final predicted result of the building operation carbon emissions.

2. The method for predicting the operating carbon emissions of cold-region buildings based on a neural network according to claim 1, wherein In step 1, the method of using the carbon emission coefficient method and using the preprocessed power consumption and heat consumption data to calculate the carbon emission data generated by the electricity consumption of the building to be tested and the carbon emission data generated by the heating heat consumption respectively is as follows: Using the formula: CE = BFS × CEI Calculate the building operation carbon emissions CE, where BFS is the building area of the test building and CEI is the building operation carbon emission intensity; The building operation carbon emission intensity is: Where EUI is the energy use index of each energy, CEF represents the carbon emission factor of each energy, and S represents the number of energies participating in the operation carbon emissions; Calculate the carbon emission data generated by the electricity consumption as: CEI 电 = EUI 电 × CEF 电 EUI 电 is the energy usage index of electric power energy, CEF 电 is the carbon emission factor of electric power energy; Calculate the carbon emission data generated by the heating heat consumption as: CEI 热 = EUI 热 × CEF 热 . EUI 热 is the energy use index for heating thermal energy, CEF 热 is the carbon emission factor for heating thermal energy.

3. The method for predicting the operating carbon emissions of cold-region buildings based on a neural network according to claim 1 or 2, characterized in that, In step 2, the method of performing a correlation analysis on the carbon emission data generated by the electricity consumption and meteorological factors to obtain the correlation coefficient is as follows: Divide the carbon emission data generated by the annual electricity consumption into 12 time periods with months as the basic unit, and perform a correlation analysis on the carbon emission data generated by the building electricity consumption and meteorological factors in each time period; Pearson product-moment correlation coefficient PCCs: Among them, cov(X,Y) is the covariance of the rank variables, and σ X σ Y is the standard deviation of the rank variables. X and Y are the carbon emission data and average temperature generated by the building's electricity consumption respectively.

4. The method for predicting the operating carbon emissions of cold-region buildings based on a neural network according to claim 3, characterized in that In step 3, the method of dividing the carbon emission data generated by the power consumption into a strong correlation group, a medium correlation group, and a weak correlation group according to the correlation coefficient is as follows: According to the absolute value of the Pearson product-moment correlation coefficient PCCs, the correlation is divided into four levels: a strong correlation group, a medium correlation group, and a weak correlation group. The weak correlation group includes weakly correlated and uncorrelated data; For the strong correlation group, the absolute value range of the correlation coefficient of the strong correlation group is 0.5 to 1.0, the absolute value range of the correlation coefficient of the medium correlation group is 0.3 to 0.5, and the absolute value range of the correlation coefficient of the weak correlation group is 0 to 0.

3.

5. The method for predicting the carbon emissions of cold-region building operations based on a neural network according to claim 3, wherein In step 3, the method of obtaining the comfortable group and the uncomfortable group is as follows: The medium correlation group is re-grouped using the outdoor weather comfort index OWCI; OWCI = C-(C ta + C tm + C h + C v + C r ) Among them, C is a constant, C ta is the average temperature term, C tm is the maximum temperature term, C h is the humidity term, C v is the wind speed term, C r is the radiation term. The data with a comfort index ≥ 7.0 are divided into the comfortable group, and the data with a comfort index ≤ 7.0 are divided into the uncomfortable group.

6. The method for predicting the carbon emissions of cold-region building operations based on a neural network according to claim 1, wherein In step 5, the long short-term neural network model includes an output gate (1), a forget gate (2), an input gate (3), a matrix calculation module (4), a first multiplier (5), a first adder (6), a second multiplier (7), and a third multiplier (8); The input data is simultaneously output to the output gate (1), the forget gate (2), the input gate (3), and the matrix calculation module (4); The input gate (3) controls the input switch using an activation function to control whether the input data is input to the third multiplier (8); The forget gate (2) uses an activation function to control whether the predicted value data of the previous time step is retained; the retained data is input to the second multiplier (7); The output gate (1) is used to control whether the input data of the current time step and the predicted value of the previous time step are converted into a hidden state, and the data that is not converted into a hidden state is transmitted to the first multiplier (5); The matrix calculation module (4) is used to perform matrix calculations on the input data to obtain a calculation result, and the calculation result is input to the third multiplier (8); The third multiplier (8) performs an element-wise matrix multiplication on the data output by the input gate (3) and the calculation result of the matrix calculation module (4) to obtain a set of candidate cell state values; the set of candidate cell state values is transmitted to the first adder (6); The second multiplier (7) performs an element-wise matrix multiplication on the result output by the first adder (6) at the previous time step and the data retained by the forget gate (2) to obtain a product result; The product result is input to the first adder (6). The first adder (6) performs an element-wise matrix addition on the product result and the set of candidate cell state values, and the result of the element-wise addition is input to the second multiplier (7) and the first multiplier (5); The first multiplier (5) performs a principal element matrix multiplication on the data that is not converted into a hidden state by the output gate (1) and the result of the element-wise addition to obtain the cell state of the next time step; The cell state of the next time step is output as the predicted value.

7. The method for predicting the carbon emissions of cold-region building operations based on a neural network according to claim 6, characterized in that, The formula for the forget gate (2) using an activation function to control whether the predicted value data of the previous time step is retained is: f t = σ(W f x t + U f h t-1 + b f ) f t represents the output of the forget gate, σ represents the activation function, with a value of 0 or 1, where 0 represents complete forgetting and 1 represents complete retention; U f represents the weight matrix of the linear transformation of the forget gate on the hidden state h t-1 at the previous time step, W f represents the input weight matrix of the forget gate, b f represents the bias of the forget gate.

8. The method for predicting the operational carbon emissions of cold-region buildings based on a neural network according to claim 7, wherein The formula for the input gate (3) using an activation function to control the input switch and control the input data is: i t = σ(W i x t + U i h t-1 + b i ) σ is the activation function, with a value of 0 or 1. 0 represents closed, and 1 represents fully open; Used to control the opening and closing of the input gate, W i Represents the input weight matrix of the input gate, U i Represents the linear transformation weight matrix of the input gate for the hidden state h at the previous time step t-1 , b i Represents the bias of the input gate, i t Represents the output of the input gate.

9. The method for predicting the carbon emissions of cold-region building operations based on a neural network according to claim 8, wherein The formula for the output gate (1) to perform a partial state transformation on the input data of the current time step and the predicted value of the previous time step is: o t = σ(W o x t + U o h t-1 + b o ) o t represents the output of the output gate, W o represents the input weight matrix of the output gate, U o represents the linear transformation weight matrix of the output gate on the hidden state h of the previous time step t-1 , b o represents the bias of the output gate.

10. The method for predicting the carbon emissions of cold-region building operations based on a neural network according to claim 9, wherein The predicted value is: h t = o t ·tanh(C t ) Among them, C t represents the cell state at time step t.

Images