Regional electric carbon emission factor prediction method and system based on multi-modal neural network

Abstract

This invention discloses a regional electric carbon emission factor prediction method and system based on a multimodal neural network, relating to the field of smart grid technology. The invention utilizes a regional electric carbon emission factor prediction model to extract the spatial correlation between inter-regional power interaction and distributed photovoltaic transmission losses, supporting hourly dynamic updates of carbon emission factors and further enhancing real-time dynamic response capabilities. It employs a hybrid attention mechanism to quantify the dual-modal characteristics of meteorology and distributed photovoltaic distribution areas, and sets dynamic contribution weights for carbon emission factors, further improving the interpretability of cross-modal causal correlations. By modeling independent branches for photovoltaic output prediction and distributed photovoltaic transmission loss prediction, the evolution paths of meteorologically sensitive and distributed photovoltaic distribution area sensitive variables are explicitly separated, avoiding error accumulation lag caused by feature coupling, improving prediction sensitivity under abrupt changes, and providing data support for the low-carbon transformation of the power system.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, and in particular to a method and system for predicting regional electrical carbon emission factors based on multimodal neural networks. Background Technology

[0002] Driven by the global energy structure's low-carbon transformation and the "dual carbon" goal, the regional electricity carbon emission factor (CEF), as a core indicator for quantifying the carbon intensity of electricity consumption, has become a key technology for accurately predicting user-side carbon footprint management, demand response optimization, and carbon market trading. Traditional CEF estimation methods mostly rely on static carbon emission factor libraries or simplified averaging models, which are difficult to adapt to the dynamic fluctuations in power system carbon intensity under high-proportion renewable energy grid integration. Especially in complex scenarios involving user-side distributed photovoltaic self-consumption, flexible electric vehicle charging, demand response strategies, and real-time electricity price interactions, traditional methods face significant bottlenecks in real-time performance, accuracy, and correlation with user behavior. With the surge in demand for low-carbon transformation on the electricity consumption side, there is an urgent need to dynamically capture the spatiotemporal heterogeneity of power system carbon intensity through data-driven methods. This will support user-side carbon emission tracing, green electricity trading matching, and energy consumption strategy optimization, providing a decision-making basis for building a new power system that coordinates "source-grid-load-storage". Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for predicting regional electrical carbon emission factors based on multimodal neural networks, so as to improve the above-mentioned technical problems.

[0004] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0005] A method for predicting regional electrical carbon emission factors based on multimodal neural networks includes:

[0006] Acquire the factors related to carbon emissions from electricity and the topology of the photovoltaic (PV) distribution area in the test region; the factors related to carbon emissions from electricity include electricity consumption data, PV distribution area data, meteorological data, distributed PV transmission loss data, and historical CEF data;

[0007] Preprocessing of the factors related to carbon emissions from electricity yields the preprocessed factors related to carbon emissions from electricity.

[0008] Construct a regional prediction model for carbon emission factors from electricity;

[0009] The topology of the photovoltaic power station area and the pre-processed factors related to electricity carbon emissions are input into the regional electricity carbon emission factor prediction model, and the regional electricity carbon emission factor prediction results are output.

[0010] Furthermore, the regional power carbon emission factor prediction model includes an output prediction sub-model, an output loss prediction sub-model, a first feature fusion layer, a dynamic carbon intensity prediction sub-model, and an FC1 layer.

[0011] The output prediction sub-model consists of a first LSTM module, a position encoding layer, a Self-Attention layer, a second feature fusion layer, and an FC2 layer, all connected in series. The first LSTM module consists of N LSTM layers connected in series. The output loss prediction sub-model consists of a second LSTM module and a GNN layer, all connected in series. The second LSTM module consists of M LSTM layers connected in series. Each LSTM layer in the first and second LSTM modules has the same structure and uses a Long Short-Term Memory artificial neural network. The GNN layer uses a graph convolutional neural network.

[0012] The dynamic carbon intensity prediction sub-model includes a cascaded encoding module, a decoding module, a linear layer, and an output layer. The encoding module includes N1 encoders cascaded in series. The decoding module includes N1 decoders cascaded in series. Each encoder includes a cascaded first hybrid attention layer and an FFN layer. The first hybrid attention layer includes a cascaded multi-attention layer and a CCA layer. Each decoder includes a cascaded masked-multi-attention layer, a second hybrid attention layer, and an FFN layer. The second hybrid attention layer includes a cascaded multi-attention layer and a CCA layer. Each FFN layer includes a cascaded input layer, a hidden block, a DSConv layer, and an output layer. The hidden block includes N2 hidden layers.

[0013] Furthermore, the training process of the regional electricity carbon emission factor prediction model includes:

[0014] Acquire the carbon emission-related factors and their labels for the training power grid, and the topology of the training photovoltaic power station area;

[0015] Meteorological training data is input into the power output prediction sub-model, and the output training prediction result is output as the power output fluctuation training feature vector.

[0016] The power consumption training data, photovoltaic substation training data, and distributed photovoltaic transmission loss training data are input into the second LSTM module, and the output power loss time series training features are obtained.

[0017] The power loss time-series training features and the training photovoltaic area topology are input into the GNN layer, and the power loss training prediction results are output as the power loss training feature tensor.

[0018] The power fluctuation training feature vector and the power loss training feature tensor are input into the first feature fusion layer, and the output is the fluctuation-loss training fusion feature.

[0019] The fluctuation-loss training fusion feature and historical CEF training data are input into the dynamic carbon intensity prediction sub-model, and the initial training prediction results of the regional electricity carbon emission factor are output.

[0020] The initial training prediction results of the regional electric carbon emission factor are input into the FC1 layer, and the training prediction results of the regional electric carbon emission factor are output.

[0021] The target loss function is calculated based on the power fluctuation training feature vector, the power loss training feature tensor, the regional electric carbon emission factor training prediction results and labels.

[0022] Based on the objective loss function, the Adam optimization algorithm is used to optimize the weight parameters of the regional electricity carbon emission factor prediction model.

[0023] Furthermore, the step of inputting meteorological training data into the power output prediction sub-model and outputting the power output training prediction result includes:

[0024] The meteorological training data is input into the first LSTM module, and the initial training features of the meteorological time series are output.

[0025] The initial meteorological time-series training features are input into the location coding layer, and the output is the meteorological time-series training features.

[0026] The meteorological time-series training features are input into the Self-Attention layer, and the corresponding key meteorological time-series training features are output.

[0027] The key features of meteorological time series training and the meteorological time series training features are input into the second feature fusion layer, and the key fusion features of meteorological time series training are output.

[0028] The key fusion features of meteorological time series training are input into the FC2 layer, and the output is the power training prediction result.

[0029] Furthermore, the step of inputting the fluctuation-loss training fusion features and historical CEF training data into the dynamic carbon intensity prediction sub-model to output the initial training prediction results of the regional electricity carbon emission factor includes:

[0030] The fluctuation-loss training fusion feature and historical CEF training data are concatenated and input into the encoding module, and the fluctuation-loss training encoded feature is output.

[0031] The fluctuation-loss training encoded features are input into the decoding module, and the fluctuation-loss training decoded features are output.

[0032] The fluctuation-loss training decoded features are input into the linear layer, and the output is the fluctuation-loss training linear features.

[0033] The fluctuation-loss training linear features are input into the output layer, and the initial training prediction results of the regional electric carbon emission factor are obtained from the output.

[0034] Furthermore, the processing procedure of the first encoder in the encoding module includes:

[0035] The fluctuation-loss training fusion feature is obtained by concatenating the fluctuation-loss training fusion feature with historical CEF training data and then performing position encoding.

[0036] The fluctuation-loss training fusion encoded features are input into the Multi-Attention layer, and the output is the fluctuation-loss training key features.

[0037] By concatenating the key features of fluctuation-loss training and the fusion-encoded features of fluctuation-loss training, the fusion-key features of fluctuation-loss training are obtained.

[0038] The key features of the fluctuation-loss training fusion are input into the CCA layer, and the output is the multimodal key features of the fluctuation-loss training.

[0039] By concatenating the key features of the multimodal fluctuation-loss training with the key features of the fusion fluctuation-loss training, the multimodal fluctuation-loss training fusion features are obtained.

[0040] The volatile-loss training multimodal fusion features are input into the FFN layer, and the output is the volatile-loss training encoded features.

[0041] Furthermore, the processing procedure of the first decoder in the decoding module includes:

[0042] The original target sequence is obtained and processed by movement and position encoding to obtain the training target sequence;

[0043] The training target sequence is input into the Masked-Multi-Attention layer, and the key features of the training target mask are output.

[0044] The training target sequence and the key features of the training target mask are concatenated to obtain the key fusion features of the training target mask;

[0045] The fluctuation-loss training encoded features and the key fusion features of the training target mask are input into the Multi-Attention layer, and the output is the fluctuation-loss training key features.

[0046] The key features of fluctuation-loss training and the key fusion features of training target mask are concatenated to obtain the key fusion features of fluctuation-loss training.

[0047] The key fusion features of fluctuation-loss training are input into the CCA layer, and the output is the multimodal key decoding features of fluctuation-loss training.

[0048] By concatenating the key decoding features of the fluctuating-loss training multimodal model and the key fusion features of the fluctuating-loss training model, the fusion decoding features of the fluctuating-loss training multimodal model are obtained.

[0049] The ripple-loss training multimodal fusion decoding features are input into the FFN layer, and the first ripple-loss training initial decoding features are output.

[0050] The initial decoding features of the first fluctuation-loss training and the multimodal fusion decoding features of the fluctuation-loss training are concatenated to obtain the fluctuation-loss training decoding features.

[0051] A regional electrical carbon emission factor prediction system based on a multimodal neural network includes:

[0052] The data acquisition module is used to acquire the factors related to carbon emissions from electricity and the topology of the photovoltaic power station area in the area to be measured; the factors related to carbon emissions from electricity include electricity consumption data, photovoltaic power station data, meteorological data, distributed photovoltaic transmission loss data, and historical CEF data;

[0053] The preprocessing module is used to preprocess the factors related to carbon emissions from electricity to obtain the preprocessed factors related to carbon emissions from electricity.

[0054] The regional electricity carbon emission factor prediction module is used to input the topology of the photovoltaic power station area and the pre-processed electricity carbon emission related factors into the regional electricity carbon emission factor prediction model to obtain the regional electricity carbon emission factor prediction results for the area.

[0055] Furthermore, the regional electricity carbon emission factor prediction module includes:

[0056] The power output prediction submodule is used to process the preprocessed meteorological data using the power output prediction sub-model to obtain the power output prediction results.

[0057] The power loss prediction submodule is used to process the pre-processed factors related to carbon emissions from electricity and the topology of the photovoltaic power station using the power loss prediction sub-model to obtain the power loss prediction results.

[0058] The dynamic carbon intensity prediction sub-model module is used to process the power output prediction results, power loss prediction results, and historical CEF data using the dynamic carbon intensity prediction sub-model to obtain the regional power carbon emission factor prediction results.

[0059] The beneficial effects of this invention are as follows:

[0060] This invention is based on a parallel feature extraction architecture of meteorological time series and distributed photovoltaic (PV) area topology. It can simultaneously analyze the time-varying law of the impact of minute-level meteorological fluctuations on PV output, as well as the spatial correlation of inter-regional power interaction on distributed PV transmission losses. It breaks through the computational bottleneck of traditional serial modeling, supports hourly dynamic updates of carbon emission factors, and further enhances real-time dynamic response capabilities.

[0061] This invention employs a hybrid attention mechanism to quantify the dual-modal characteristics of meteorological and distributed photovoltaic power station areas, and sets dynamic contribution weights for carbon emission factors. It establishes a visualized analysis of the causal transmission chain of "irradiance fluctuation → power output change → loss adjustment → carbon emission response", overcoming the problem of delayed correction of prediction results caused by the opaque mechanism of traditional black box models, and further improving the interpretability of cross-modal causal associations.

[0062] This invention uses independent branch modeling for photovoltaic power output prediction and distributed photovoltaic transmission loss prediction to explicitly separate the evolution paths of weather-sensitive and distributed photovoltaic area-sensitive variables, avoid error accumulation lag caused by feature coupling, improve prediction sensitivity under sudden operating conditions, and provide data support for the low-carbon transformation of the power system. Attached Figure Description

[0063] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart of the method in an embodiment of the present invention;

[0065] Figure 2 This is a structural diagram of the regional electric carbon emission factor prediction model in an embodiment of the present invention;

[0066] Figure 3 This is a structural diagram of the output prediction sub-model in an embodiment of the present invention;

[0067] Figure 4 This is a structural diagram of the decoder and encoder in an embodiment of the present invention;

[0068] Figure 5 This is a structural diagram of the regional electric carbon emission factor prediction system in an embodiment of the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0070] Please see Figure 1 This embodiment provides a method for predicting regional electrical carbon emission factors based on multimodal neural networks, which includes:

[0071] S1. Obtain the relevant factors of carbon emissions from electricity and the topology of the photovoltaic power station area in the area to be tested; the relevant factors of carbon emissions from electricity include electricity consumption data, photovoltaic power station area data, meteorological data, distributed photovoltaic transmission loss data and historical CEF data.

[0072] Electricity consumption data includes electricity consumption from various energy sources, such as coal-fired power, natural gas, wind power, and photovoltaic power. Photovoltaic distribution data includes output data, installed capacity, and geographical location of distributed photovoltaic systems. Meteorological data includes solar radiation intensity, temperature, and cloud cover (which affects photovoltaic output). Distributed photovoltaic transmission loss data represents the regional distributed photovoltaic transmission loss rate. Historical CEF data represents the regional historical carbon emission factor. CEF is the regional electricity carbon emission factor.

[0073] S2. Preprocess the factors related to carbon emissions from electricity to obtain the preprocessed factors related to carbon emissions from electricity.

[0074] S2 includes:

[0075] S2-1. Normalize the electricity consumption data and meteorological data to obtain normalized electricity consumption data and normalized meteorological data. Before normalization, data cleaning and data filling operations can also be performed.

[0076] S2-2. Time alignment is performed on photovoltaic power station data, meteorological data, distributed photovoltaic transmission loss data, historical CEF data, electricity consumption normalization data, and meteorological normalization data, unifying the time granularity to the hour level, to obtain the corresponding time-aligned photovoltaic power station data, meteorological data, distributed photovoltaic transmission loss data, historical CEF data, and electricity consumption data, which are the preprocessed factors related to carbon emissions from electricity.

[0077] S3. Construct a regional electricity carbon emission factor prediction model;

[0078] like Figure 2As shown, the regional power carbon emission factor prediction model includes an output prediction sub-model, an output loss prediction sub-model, a first feature fusion layer, a dynamic carbon intensity prediction sub-model, and an FC1 layer (fully connected layer).

[0079] like Figure 3 As shown, the output prediction sub-model includes a first LSTM module, a position encoding layer, a Self-Attention layer, a second feature fusion layer, and an FC2 layer (fully connected layer) connected in series. The first LSTM module consists of N LSTM layers connected in series. The output loss prediction sub-model includes a second LSTM module and a GNN layer connected in series. The second LSTM module consists of M LSTM layers connected in series. Each LSTM layer in the first and second LSTM modules has the same structure and uses a Long Short-Term Memory (LSTM) artificial neural network. The position encoding layer uses an MLP (Multilayer Perceptron) and the GNN layer uses a graph convolutional neural network. In this embodiment, the values ​​of N and M are 3 and 6, respectively. The hidden layer dimension of the LSTM layer is 128, and the number of self-attention heads in the Self-Attention layer is 3.

[0080] The dynamic carbon intensity prediction sub-model includes a cascaded encoding module, a decoding module, a linear layer, and an output layer; the encoding module consists of N1 encoders cascaded in series; the decoding module consists of N1 decoders cascaded in series; such as Figure 4 As shown, each encoder includes a first hybrid attention layer and an FFN layer (feedforward neural network) in series; the first hybrid attention layer includes a multi-attention layer and a CCA layer (cross-modal cross-attention mechanism) in series; each decoder includes a masked-multi-attention layer, a second hybrid attention layer, and an FFN layer in series; the second hybrid attention layer includes a multi-attention layer and a CCA layer in series; each FFN layer includes an input layer, a hidden block, a DSConv layer (depthseparable convolutional layer), and an output layer in series; the hidden block includes N2 hidden layers; the output layer uses the Softmax function. In this embodiment, N1 is set to 3.

[0081] The training process of the regional electricity carbon emission factor prediction model includes:

[0082] S3-1. Obtain the relevant factors for training electricity carbon emissions and their labels, and the training photovoltaic (PV) substation topology G = (V, E); where V represents the nodes in the PV substation topology, i.e., the PV substations; and E represents the edges in the PV substation topology, i.e., the transmission lines between PV substations. The relevant factors for training electricity carbon emissions include electricity consumption training data, PV substation training data, meteorological training data, distributed PV transmission loss training data, and historical CEF training data.

[0083] S3-2. Input the meteorological training data into the power output prediction sub-model, output the power output training prediction result, and use it as the power output fluctuation training feature vector.

[0084] S3-2 includes:

[0085] S3-2-1. Input the meteorological training data into the first LSTM module to capture long-term trends such as weather cycles and output the initial training features of the meteorological time series.

[0086] In this embodiment, N is 3, therefore, S3-2-1 includes:

[0087] The meteorological training data is input into the first LSTM layer of the first LSTM module, and the first meteorological time-series training feature is output. The corresponding formula is:

[0088] it=σ(Wxixt+Whiht-1+bi);

[0089] ft=σ(Wxfxt+Whfht-1+bf);

[0090] gt=tanh(Wxgxt+Whght-1+bg);

[0091] ft=σ(Wxoxt+Whoht-1+bo);

[0092] ct = ft·ct-1 + it·gt;

[0093] ht = ot·tanh(ct);

[0094]

[0095] Where bi, bf, bg, and bo represent the bias terms of the input gate, forget gate, candidate state, and output gate, respectively; σ(·) and tanh(·) represent the activation function and hyperbolic tangent function, respectively; xt represents the input data at time step t; Wxi and Whi represent the matrix weights of the input gate; ht-1 and ht represent the hidden states at time steps t-1 and t, respectively; it represents the output data of the input gate at time step t; Wxf and Whf both represent the matrix weights of the forget gate; ft represents the output data of the forget gate at time step t; and gt represents the candidate state at time step t. The values ​​are: Wxg and Whg represent the matrix weights of the candidate states, Wxo and Who represent the matrix weights of the output gates, ct and ct-1 represent the cell states at time steps t and t-1, respectively, ot represents the output data of the output gate at time step t, HLSTM,1 represents the output data of the first LSTM layer in the first LSTM module, i.e., the first meteorological time series training feature, h1, h2, and hT represent the hidden states at the first, second, and last time steps, respectively, R represents a constant, and T and dh represent the total number of time steps and the dimension of the hidden layer, respectively.

[0096] The first meteorological time series training feature is input into the second LSTM layer in the first LSTM module, and the second meteorological time series training feature is output.

[0097] The second meteorological time series training features are input into the third LSTM layer in the first LSTM module, and the initial meteorological time series training features are output.

[0098] S3-2-2: Input the initial meteorological time series training features into the location coding layer (MLP multilayer perceptron) to enhance the output prediction sub-model's perception of the time series order, avoid the loss of location information in long sequences, and output the meteorological time series training features.

[0099] S3-2-3. Input the meteorological time series training features into the Self-Attention layer and output the corresponding key meteorological time series training features. Introduce an adaptive weight mechanism in the Self-Attention layer to automatically allocate feature weights and focus on key time step information.

[0100] Therefore, the formula corresponding to the Self-Attention layer is:

[0101]

[0102] Where dk represents the vector dimension, Q, K, and V represent the query matrix, key matrix, and value matrix, respectively, WQ, WK, and WV represent the weight parameters corresponding to the query matrix Q, key matrix K, and value matrix V, respectively, Attention(Q,K,V) represents the output of the Self-Attention layer (key features for meteorological time series training), softmax(·) represents the activation function, KT represents the transpose of the key matrix, HLSTM,3 represents the output data of the third LSTM layer in the first LSTM module, i.e., the initial training features for meteorological time series, and MLP(·) represents the multilayer perceptron.

[0103] S3-2-4. Input the key features of meteorological time series training and the meteorological time series training features into the second feature fusion layer, and output the key fusion features of meteorological time series training.

[0104] S3-2-5. Input the key fusion features of meteorological time series training into the FC2 layer and output the power training prediction results.

[0105] The formulas corresponding to S3-2-4 and S3-2-5 are:

[0106] Hfusion=MLP(HLSTM,3)+Attention(Q,K,V);

[0107]

[0108] Where Hfusion represents the key fusion features for meteorological time series training, dh represents the dimension of the hidden state at each time step, hfinal represents the hidden state at the last time step, Wfc and bfc represent the weight matrix and bias term of the FC2 layer, respectively, and L represents the output dimension of the fully connected layer. This indicates the predicted results of the force training.

[0109] This invention employs a combination of a first LSTM module, a position encoding layer, and a self-attention mechanism. This approach can capture long-term trends such as weather cycles, deeply perceive time series, adaptively adjust feature weights, improve the model's robustness to outliers, and accurately predict photovoltaic power output fluctuations, providing a reliable basis for grid dispatching and energy storage management.

[0110] S3-3. Input the electricity consumption training data, photovoltaic substation training data, and distributed photovoltaic transmission loss training data into the second LSTM module to capture complex nonlinear relationships and output the power loss time series training features to improve the estimation accuracy of the power prediction sub-model.

[0111] In this embodiment, the value of M is 6, therefore, S3-3 includes:

[0112] S3-3-1. After splicing the power consumption training data, photovoltaic substation training data, and distributed photovoltaic transmission loss training data, input them into the first LSTM layer in the second LSTM module to obtain the first output loss time series training feature.

[0113] S3-3-2. Input the first output loss time-series training feature into the second LSTM layer in the second LSTM module to obtain the second output loss time-series training feature.

[0114] S3-3-3: Input the second output loss time-series training feature into the third LSTM layer in the second LSTM module to obtain the third output loss time-series training feature.

[0115] S3-3-4. Input the third output loss time-series training feature into the fourth LSTM layer in the second LSTM module to obtain the fourth output loss time-series training feature.

[0116] S3-3-5. Input the fourth output loss time-series training feature into the fifth LSTM layer in the second LSTM module to obtain the fifth output loss time-series training feature.

[0117] S3-3-6. Input the fifth output loss time-series training feature into the sixth LSTM layer in the second LSTM module to obtain the output loss time-series training feature.

[0118] S3-4. Input the power loss time-series training features and the training photovoltaic power station topology into the GNN layer, and output the power loss training prediction results, which are used as the power loss training feature tensor. The corresponding process is as follows:

[0119] S3-4-1. Use the time-series training features of output loss as node features;

[0120] S3-4-2. Use the characteristics (impedance, capacity) of the transmission lines in the training photovoltaic area topology as edge features;

[0121] S3-4-3. Treat the topology of the training photovoltaic area as a graph structure and define the connection relationships between nodes and edges;

[0122] S3-4-4: Input the node features, edge features and graph structure into the graph convolutional neural network, and output the spatiotemporal features of the power loss training.

[0123] S3-4-5. Input the spatiotemporal features of the power loss training into the fully connected layer, and output the power loss training prediction results.

[0124] Therefore, the formula corresponding to S3-4 is:

[0125]

[0126] Where Concat(·) represents the concatenation function, AGG(·) represents the aggregation function, hv(l+1) and hv(l) represent the embeddings of node v in the (l+1)th and 1st layers, respectively, μ represents a node, hμ(l) represents the embedding of node μ in the 1st layer, W(l) represents the learnable weight matrix in the 1st layer, and N(v) represents the set of neighbors of node v. denoted as the output loss training prediction result, Wfc and bfc represent the weight matrix and bias term of the fully connected layer, respectively, and hT represents the spatiotemporal features of the output loss training.

[0127] S3-5. Input the power fluctuation training feature vector and the power loss training feature tensor into the first feature fusion layer, and output the fluctuation-loss training fusion feature.

[0128] S3-6. Input the fluctuation-loss training fusion features and historical CEF training data into the dynamic carbon intensity prediction sub-model, and output the initial training prediction results of the regional electricity carbon emission factor.

[0129] S3-6 includes:

[0130] S3-6-1. After concatenating the fluctuation-loss training fusion features and historical CEF training data, input them into the encoding module and output the fluctuation-loss training encoded features.

[0131] In this embodiment, N1 is 3. Therefore, taking the first encoder in the encoding module as an example, the corresponding process includes:

[0132] S3-6-1-1. After concatenating the fluctuation-loss training fusion feature with the preprocessed historical CEF training data, position encoding is performed to obtain the fluctuation-loss training fusion encoded feature.

[0133] S3-6-1-2. Input the fluctuation-loss training fusion encoded features into the Multi-Attention layer to obtain the fluctuation-loss training key features;

[0134] S3-6-1-3. Concatenate the key features of fluctuation-loss training and the fusion coding features of fluctuation-loss training to obtain the fusion key features of fluctuation-loss training.

[0135] S3-6-1-4. Input the key features of the fluctuation-loss training fusion into the CCA layer and output the multimodal key features of the fluctuation-loss training.

[0136] S3-6-1-5. Concatenate the key features of the fluctuation-loss training multimodal model with the key features of the fluctuation-loss training fusion to obtain the fluctuation-loss training multimodal fusion features.

[0137] S3-6-1-6. Input the fluctuation-loss training multimodal fusion features into the FFN layer and output the fluctuation-loss training encoded features.

[0138] The formula corresponding to S3-6-1-6 is:

[0139] hi = ReLU(Wihi-1 + bi);

[0140] hdepth=DepthwiseConv(hhidden);

[0141] hpoint=DepthwiseConv(hdepth);

[0142] Y=softmax(Wouthpoint+bout);

[0143] Where Wi and bi represent the weights and biases of the i-th hidden layer, hi and hi-1 represent the outputs of the i-th and (i-1)-th hidden layers, respectively; Wout and bout represent the weights and biases of the output layer, respectively; ReLU(·) represents the activation function; DepthwiseConv(·) represents the depthwise convolution operation; PointwiseConv(·) represents the pointwise convolution operation; hhidden represents the output of the last hidden layer; hdepth represents the output of the depthwise convolution operation; hpoint represents the output of the pointwise convolution operation; and Y represents the fluctuating-loss training encoded features.

[0144] The processing procedures for the remaining encoders are the same as those in S3-6-1-2 to S3-6-1-6. Except for the first encoder, whose input data is the concatenation of fluctuation-loss training fusion features and preprocessed historical CEF training data, the input data for the remaining encoders is the output data of the previous encoder, and the output data of the last encoder is one input data of the first decoder.

[0145] This invention employs a hybrid attention mechanism, utilizing multi-head attention to capture features of input data from different perspectives, enhancing the model's expressive and generalization abilities, and demonstrating excellent performance in the initial processing of complex sequence data. A cross-modal attention mechanism is introduced to effectively capture the correlations between different modal features (such as meteorological data, photovoltaic power output data, historical carbon emission data, etc.), improving the model's ability to model complex coupling relationships. The contribution weights of different modal features are dynamically adjusted to avoid error accumulation caused by feature coupling, thereby improving the accuracy of feature fusion, quantifying the dynamic contribution weights of different modal features to carbon emission factors, and establishing a visual analysis of the causal transmission chain. The introduction of depthwise separable convolutional layers in the FFN significantly reduces computational cost and the number of model parameters, improving computational efficiency, reducing the risk of overfitting, and accelerating the model's training and inference processes, significantly improving the accuracy and efficiency of the carbon emission prediction model.

[0146] S3-6-2. Input the fluctuation-loss training encoded features into the decoding module and output the fluctuation-loss training decoded features.

[0147] Taking the first decoder in the decoding module as an example, the process is explained as follows:

[0148] S3-6-2-1. Obtain the original target sequence and perform shifting (left shift by one position) and position encoding to obtain the training target sequence. During the prediction process, the input data for other decoders is the output of the previous decoder. The input of the first decoder includes the output of the last encoder and the target sequence. The current target sequence is the output sequence gradually generated by the dynamic carbon intensity prediction sub-model.

[0149] S3-6-2-2: Input the training target sequence into the Masked-Multi-Attention layer, and output the key features of the training target mask.

[0150] S3-6-2-3. Concatenate the training target sequence and the key features of the training target mask to obtain the key fusion features of the training target mask;

[0151] S3-6-2-4. Input the fluctuation-loss training encoded features and the key fusion features of the training target mask into the Multi-Attention layer, and output the fluctuation-loss training key features.

[0152] S3-6-2-5. Concatenate the key features of fluctuation-loss training and the key fusion features of training target mask to obtain the key fusion features of fluctuation-loss training.

[0153] S3-6-2-6: Input the key fusion features of fluctuation-loss training into the CCA layer, and output the key decoding features of fluctuation-loss training multimodal;

[0154] S3-6-2-7. The key decoding features of the fluctuation-loss training multimodal model and the key fusion features of the fluctuation-loss training model are concatenated to obtain the fusion decoding features of the fluctuation-loss training multimodal model.

[0155] S3-6-2-8. Input the ripple-loss training multimodal fusion decoding features into the FFN layer and output the first ripple-loss training initial decoding features; the processing of the FFN layer in the decoder is the same as that in the encoder.

[0156] S3-6-2-9. The initial decoding features of the first fluctuation-loss training and the multimodal fusion decoding features of the fluctuation-loss training are concatenated to obtain the fluctuation-loss training decoding features.

[0157] S3-6-3. Input the fluctuation-loss training decoding features into the linear layer and output the fluctuation-loss training linear features.

[0158] S3-6-4. Input the fluctuation-loss training linear features into the output layer to obtain the initial training prediction results of the regional electric carbon emission factor.

[0159] S3-7. Input the initial training prediction results of the regional electric carbon emission factor into the FC1 layer and output the training prediction results of the regional electric carbon emission factor.

[0160] S3-8. Based on the power output fluctuation training feature vector, the power loss training feature tensor, the regional electric carbon emission factor training prediction results and labels, calculate the target loss function;

[0161] The target loss function includes a first loss function, a second loss function, and a third loss function. The first loss function, L1, is the loss function corresponding to the output prediction sub-model, which includes the MSE loss function LMSE and the fluctuation smoothing regularization loss function Lsmooth. The second loss function, L2, is the loss function corresponding to the output loss prediction sub-model. The third loss function, L3, is the loss function corresponding to the dynamic carbon intensity prediction sub-model, which includes the main network loss Lmain and the L1 regularization loss function LL1. Due to the introduction of a hybrid attention mechanism (self-attention mechanism + cross-modal cross-attention mechanism) and the design of multiple prediction models to construct the regional electricity carbon emission factor prediction model, a large number of parameters will be adjusted during training. This makes the regional electricity carbon emission factor prediction model more prone to overfitting, especially when there is noise or imbalance in the multimodal data. Therefore, the fluctuation smoothing regularization loss function Lsmooth and the L1 regularization loss function LL1 are introduced to reduce the overfitting risk of the regional electricity carbon emission factor prediction model and increase its stability.

[0162] Therefore, the formula corresponding to the target loss function Ltotal is:

[0163] Ltotal = αL1 + βL2 + λ1L3;

[0164] L1 = LMSE + Lsmooth;

[0165]

[0166] L3 = Lmain + LL1;

[0167]

[0168] Where α, β, λ1, and λ2 represent coefficients, T represents the number of power fluctuation training feature vectors, and ∑(·) represents the summation function. Let represent the t-th power fluctuation training feature vector and the (t-1)-th power fluctuation training feature vector, respectively. N² represents the number of power loss training feature tensors. yi, Ypre,i and Ytrue,i represent the training feature tensor of the i-th power loss and its corresponding label (power loss result), respectively. Ypre,i and Ytrue,i represent the training prediction result of the i-th regional carbon emission factor and its corresponding label (actual result of the regional carbon emission factor), respectively.

[0169] S3-9. Based on the objective loss function, the Adam optimization algorithm is used to optimize the weight parameters of the regional electricity carbon emission factor prediction model.

[0170] S4. Input the photovoltaic power station topology and pre-processed electricity carbon emission related factors into the regional electricity carbon emission factor prediction model, and output the regional electricity carbon emission factor prediction results for the region. The regional electricity carbon emission factor prediction results are hourly dynamic prediction values, i.e., hourly carbon emissions.

[0171] like Figure 5 As shown, a regional electrical carbon emission factor prediction system based on a multimodal neural network includes:

[0172] The data acquisition module is used to acquire the factors related to carbon emissions from electricity and the topology of the photovoltaic power station area in the area to be measured; the factors related to carbon emissions from electricity include electricity consumption data, photovoltaic power station data, meteorological data, distributed photovoltaic transmission loss data, and historical CEF data;

[0173] The preprocessing module is used to preprocess the factors related to carbon emissions from electricity to obtain the preprocessed factors related to carbon emissions from electricity.

[0174] The regional electricity carbon emission factor prediction module is used to input the topology of the photovoltaic power station area and the pre-processed electricity carbon emission related factors into the regional electricity carbon emission factor prediction model to obtain the regional electricity carbon emission factor prediction results for the area.

[0175] The regional electricity carbon emission factor prediction module includes:

[0176] The power output prediction submodule is used to process the preprocessed meteorological data using the power output prediction sub-model to obtain the power output prediction results.

[0177] The power loss prediction submodule is used to process the pre-processed factors related to carbon emissions from electricity and the topology of the photovoltaic power station using the power loss prediction sub-model to obtain the power loss prediction results.

[0178] The dynamic carbon intensity prediction sub-model module is used to process the power output prediction results, power loss prediction results, and historical CEF data using the dynamic carbon intensity prediction sub-model to obtain the regional power carbon emission factor prediction results.

[0179] In summary, this invention constructs a multimodal neural network model that effectively integrates multi-source heterogeneous data related to electricity carbon emissions, including electricity consumption data, photovoltaic (PV) distribution area data, meteorological data, distributed PV transmission loss data, and historical CEF data. Employing advanced algorithms such as LSTM, self-attention mechanisms, GNN, and Transformer, it achieves accurate predictions of PV output fluctuations, transmission losses, and carbon emission intensity. The scheme possesses real-time dynamic response capabilities, quickly analyzing the impact of meteorological fluctuations and power interaction on PV output and losses, overcoming the limitations of traditional sequential modeling. Through cross-modal causal correlation modeling, the dynamic contribution of different modal features to carbon emission factors is quantified, improving the model's interpretability. Decoupling modeling of complex coupling relationships effectively separates meteorological-sensitive and PV distribution area-sensitive variables, avoiding error accumulation caused by feature coupling. Furthermore, the scheme introduces fluctuation smoothing regularization and L1 regularization loss functions, reducing the risk of model overfitting and increasing stability. This scheme significantly reduces computational complexity while improving prediction accuracy, enhancing model efficiency and robustness, and providing strong data support for the low-carbon transformation of the power system.

[0180] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting regional electrical carbon emission factors based on multimodal neural networks, characterized in that, include: Acquire the factors related to carbon emissions from electricity and the topology of the photovoltaic (PV) distribution area in the test region; the factors related to carbon emissions from electricity include electricity consumption data, PV distribution area data, meteorological data, distributed PV transmission loss data, and historical CEF data; Preprocessing of the factors related to carbon emissions from electricity yields the preprocessed factors related to carbon emissions from electricity. Construct a regional prediction model for carbon emission factors from electricity; The topology of the photovoltaic power station area and the pre-processed factors related to electricity carbon emissions are input into the regional electricity carbon emission factor prediction model, and the regional electricity carbon emission factor prediction results for the region are output. The regional electricity carbon emission factor prediction model includes an output prediction sub-model, an output loss prediction sub-model, a first feature fusion layer, a dynamic carbon intensity prediction sub-model, and an FC1 layer. The output prediction sub-model consists of a first LSTM module, a position encoding layer, a Self-Attention layer, a second feature fusion layer, and an FC2 layer, all connected in series. The first LSTM module consists of N LSTM layers connected in series. The output loss prediction sub-model consists of a second LSTM module and a GNN layer, all connected in series. The second LSTM module consists of M LSTM layers connected in series. Each LSTM layer in the first and second LSTM modules has the same structure and uses a Long Short-Term Memory artificial neural network. The GNN layer uses a graph convolutional neural network. The dynamic carbon intensity prediction sub-model includes a cascaded encoding module, a decoding module, a linear layer, and an output layer. The encoding module includes N1 encoders cascaded in series. The decoding module includes N1 decoders cascaded in series. Each encoder includes a cascaded first hybrid attention layer and an FFN layer. The first hybrid attention layer includes a cascaded multi-attention layer and a CCA layer. Each decoder includes a cascaded masked-multi-attention layer, a second hybrid attention layer, and an FFN layer. The second hybrid attention layer includes a cascaded multi-attention layer and a CCA layer. Each FFN layer includes a cascaded input layer, a hidden block, a DSConv layer, and an output layer. The hidden block includes N2 hidden layers.

2. The regional electrical carbon emission factor prediction method based on multimodal neural networks according to claim 1, characterized in that, The training process of the regional electricity carbon emission factor prediction model includes: Acquire the carbon emission-related factors and their labels for the training power grid, and the topology of the training photovoltaic power station area; Meteorological training data is input into the power output prediction sub-model, and the output training prediction result is output as the power output fluctuation training feature vector. The power consumption training data, photovoltaic substation training data, and distributed photovoltaic transmission loss training data are input into the second LSTM module, and the output power loss time series training features are obtained. The power loss time-series training features and the training photovoltaic area topology are input into the GNN layer, and the power loss training prediction results are output as the power loss training feature tensor. The power fluctuation training feature vector and the power loss training feature tensor are input into the first feature fusion layer, and the output is the fluctuation-loss training fusion feature. The fluctuation-loss training fusion feature and historical CEF training data are input into the dynamic carbon intensity prediction sub-model, and the initial training prediction results of the regional electricity carbon emission factor are output. The initial training prediction results of the regional electric carbon emission factor are input into the FC1 layer, and the training prediction results of the regional electric carbon emission factor are output. The target loss function is calculated based on the power fluctuation training feature vector, the power loss training feature tensor, the regional electric carbon emission factor training prediction results and labels. Based on the objective loss function, the Adam optimization algorithm is used to optimize the weight parameters of the regional electricity carbon emission factor prediction model.

3. The regional electrical carbon emission factor prediction method based on multimodal neural networks according to claim 2, characterized in that, The process of inputting meteorological training data into the power output prediction sub-model and outputting power output training prediction results includes: The meteorological training data is input into the first LSTM module, and the initial training features of the meteorological time series are output. The initial meteorological time-series training features are input into the location coding layer, and the output is the meteorological time-series training features. The meteorological time-series training features are input into the Self-Attention layer, and the corresponding key meteorological time-series training features are output. The key features of meteorological time series training and the meteorological time series training features are input into the second feature fusion layer, and the key fusion features of meteorological time series training are output. The key fusion features of meteorological time series training are input into the FC2 layer, and the output is the power training prediction result.

4. The regional electrical carbon emission factor prediction method based on multimodal neural networks according to claim 2, characterized in that, The process of inputting the fluctuation-loss training fusion features and historical CEF training data into the dynamic carbon intensity prediction sub-model, and outputting the initial training prediction results of the regional electricity carbon emission factor, includes: The fluctuation-loss training fusion feature and historical CEF training data are concatenated and input into the encoding module, and the fluctuation-loss training encoded feature is output. The fluctuation-loss training encoded features are input into the decoding module, and the fluctuation-loss training decoded features are output. The fluctuation-loss training decoded features are input into the linear layer, and the output is the fluctuation-loss training linear features. The fluctuation-loss training linear features are input into the output layer, and the initial training prediction results of the regional electric carbon emission factor are obtained from the output.

5. The regional electrical carbon emission factor prediction method based on multimodal neural networks according to claim 4, characterized in that, The processing procedure of the first encoder in the encoding module includes: The fluctuation-loss training fusion feature is obtained by concatenating the fluctuation-loss training fusion feature with historical CEF training data and then performing position encoding. The fluctuation-loss training fusion encoded features are input into the Multi-Attention layer, and the output is the fluctuation-loss training key features. By concatenating the key features of fluctuation-loss training and the fusion-encoded features of fluctuation-loss training, the fusion-key features of fluctuation-loss training are obtained. The key features of the fluctuation-loss training fusion are input into the CCA layer, and the output is the multimodal key features of the fluctuation-loss training. By concatenating the key features of the multimodal fluctuation-loss training with the key features of the fusion fluctuation-loss training, the multimodal fluctuation-loss training fusion features are obtained. The volatile-loss training multimodal fusion features are input into the FFN layer, and the output is the volatile-loss training encoded features.

6. The regional electrical carbon emission factor prediction method based on multimodal neural networks according to claim 4, characterized in that, The processing procedure of the first decoder in the decoding module includes: The original target sequence is obtained and processed by movement and position encoding to obtain the training target sequence; The training target sequence is input into the Masked-Multi-Attention layer, and the key features of the training target mask are output. The training target sequence and the key features of the training target mask are concatenated to obtain the key fusion features of the training target mask; The fluctuation-loss training encoded features and the key fusion features of the training target mask are input into the Multi-Attention layer, and the output is the fluctuation-loss training key features. The key features of fluctuation-loss training and the key fusion features of training target mask are concatenated to obtain the key fusion features of fluctuation-loss training. The key fusion features of fluctuation-loss training are input into the CCA layer, and the output is the multimodal key decoding features of fluctuation-loss training. By concatenating the key decoding features of the fluctuating-loss training multimodal model and the key fusion features of the fluctuating-loss training multimodal model, the fusion decoding features of the fluctuating-loss training multimodal model are obtained. The ripple-loss training multimodal fusion decoding features are input into the FFN layer, and the first ripple-loss training initial decoding features are output. The initial decoding features of the first fluctuation-loss training and the multimodal fusion decoding features of the fluctuation-loss training are concatenated to obtain the fluctuation-loss training decoding features.

7. A regional electrical carbon emission factor prediction system based on a multimodal neural network, used to implement the regional electrical carbon emission factor prediction method based on a multimodal neural network as described in any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to acquire the factors related to carbon emissions from electricity and the topology of the photovoltaic power station area in the area to be measured; the factors related to carbon emissions from electricity include electricity consumption data, photovoltaic power station data, meteorological data, distributed photovoltaic transmission loss data, and historical CEF data; The preprocessing module is used to preprocess the factors related to carbon emissions from electricity to obtain the preprocessed factors related to carbon emissions from electricity. The regional electricity carbon emission factor prediction module is used to input the topology of the photovoltaic power station area and the pre-processed electricity carbon emission related factors into the regional electricity carbon emission factor prediction model to obtain the regional electricity carbon emission factor prediction results for the region. The regional electricity carbon emission factor prediction model includes an output prediction sub-model, an output loss prediction sub-model, a first feature fusion layer, a dynamic carbon intensity prediction sub-model, and an FC1 layer. The output prediction sub-model consists of a first LSTM module, a position encoding layer, a Self-Attention layer, a second feature fusion layer, and an FC2 layer, all connected in series. The first LSTM module consists of N LSTM layers connected in series. The output loss prediction sub-model consists of a second LSTM module and a GNN layer, all connected in series. The second LSTM module consists of M LSTM layers connected in series. Each LSTM layer in the first and second LSTM modules has the same structure and uses a Long Short-Term Memory artificial neural network. The GNN layer uses a graph convolutional neural network. The dynamic carbon intensity prediction sub-model includes a cascaded encoding module, a decoding module, a linear layer, and an output layer. The encoding module includes N1 encoders cascaded in series. The decoding module includes N1 decoders cascaded in series. Each encoder includes a cascaded first hybrid attention layer and an FFN layer. The first hybrid attention layer includes a cascaded multi-attention layer and a CCA layer. Each decoder includes a cascaded masked-multi-attention layer, a second hybrid attention layer, and an FFN layer. The second hybrid attention layer includes a cascaded multi-attention layer and a CCA layer. Each FFN layer includes a cascaded input layer, a hidden block, a DSConv layer, and an output layer. The hidden block includes N2 hidden layers.

8. The regional electrical carbon emission factor prediction system based on a multimodal neural network according to claim 7, characterized in that, The regional electricity carbon emission factor prediction module includes: The power output prediction submodule is used to process the preprocessed meteorological data using the power output prediction sub-model to obtain the power output prediction results. The power loss prediction submodule is used to process the pre-processed factors related to carbon emissions from electricity and the topology of the photovoltaic power station using the power loss prediction sub-model to obtain the power loss prediction results. The dynamic carbon intensity prediction sub-model module is used to process the power output prediction results, power loss prediction results, and historical CEF data using the dynamic carbon intensity prediction sub-model to obtain the regional power carbon emission factor prediction results.

Images