CCS rapid simulation method based on multi-feature fusion neural network

Abstract

The invention relates to the technical field of carbon capture and sequestration (CCS) simulation, and particularly provides a CCS rapid simulation method based on a multi-feature fusion neural network. According to the method, a U-FNO network fusing a U neural network (U-Net) and a Fourier neural operator (FNO) is constructed, and the prediction precision of a pressure field and a saturation field on a long-time sequence in the CO2 geological sequestration process is improved; meanwhile, a yield sequence prediction network based on a transformer encoder (TE) is adopted, the output end of the U-FNO network is directly connected to the input end of a TE network, and direct prediction from geological parameters and well control curves to pressure and saturation field time sequence distribution and corresponding yield curves is achieved; the method effectively overcomes the problem that the existing CCS simulation technology is difficult to process multi-feature input and output at the same time, improves the prediction efficiency of the influence of the CO2 storage rate on the reservoir safety, and can provide an efficient and reliable intelligent simulation tool for CCS engineering optimization and history fitting.

Description

Technical Field

[0001] The present invention relates to the technical field of carbon capture and storage (CCS) simulation, and specifically provides a CCS fast simulation method based on a multi-feature fusion neural network. Background Art

[0002] CO2 capture and storage (CCS) has become one of the key strategies for global climate change mitigation and significant reduction of greenhouse gas emissions. Saline aquifers, as reservoirs for geological storage of CO2, have great storage potential. There are three main challenges in developing efficient saline aquifer CO2 storage technology: assessing the storage capacity of saline aquifers to maximize CO2 injection, simulating injection rates (injectability) and well locations under geological uncertainty parameters, and assessing geological risks during CO2 injection. In order to apply carbon storage technology on a large scale in saline aquifers and depleted oil and gas reservoirs, CO2 injection methods must be technically feasible, safe, and economically reasonable, and repeatedly simulated for different storage strategies.

[0003] During the CO2 injection process, excessive pressure accumulation will bring geological risks. In order to prevent the geological risk of rock fracture due to the continuous increase in pressure, the deployment of production wells can effectively alleviate the pressure accumulation, thereby increasing the amount of CO2 that can be injected. However, starting the production well too early may cause CO2 to break through the production well, resulting in inefficient CO2 injection cycle. Therefore, it is crucial to quickly simulate and predict the impact of different CO2 injection rates on the occurrence of related geological risks in the geological prediction well injection system and the brine well opening time.

[0004] With the advancement of artificial intelligence technology, machine learning and deep learning have promoted the development of proxy models in CCS and environmental science research, but the existing CCS simulation methods are difficult to effectively handle multi-feature inputs and outputs, which makes it impossible to efficiently predict the impact of CO2 storage rate on reservoir safety. Summary of the invention

[0005] The purpose of the present invention is to provide a CCS rapid simulation method based on a multi-feature fusion neural network, aiming to solve the technical problem that the CCS simulation method in the prior art is difficult to effectively handle multi-feature input and output, which leads to the inability to efficiently predict the impact of CO2 storage rate on reservoir safety.

[0006] To achieve the above purpose, the present invention adopts a CCS fast simulation method based on multi-feature fusion neural network.

[0007] The steps include:

[0008] Establish a U-FNO architecture suitable for CCS fast simulation and direct output of pressure and saturation status in the time dimension;

[0009] Based on the U-FNO architecture, a Transformer module with self-attention mechanism is introduced to obtain a fast simulation alternative model for building CCS;

[0010] Permeability and well control sequences were selected as model inputs, and SGeMS was used to perform sequential Gaussian simulation to randomly generate 10 permeability samples and 500 sets of well control sequences for model training;

[0011] Finally, an overall performance analysis method is proposed, an error measurement mechanism is introduced, and the simulation is completed.

[0012] Among them, when constructing the alternative model:

[0013] In the prediction of CO2 reservoir storage process under geological uncertainty, the well rate of injection and production wells is adjusted in each reservoir model at each control time step; the mapping relationship required to be achieved by the relevant alternative model is expressed as:

[0014] X=[P,S,q i ,q p ]=g(m,w i ,w p )

[0015] The geological model is represented by m, which is An element of N, where nb is defined as nx×ny×nz, which represents the total number of grid blocks in the geological model, and nx, ny, and nz correspond to the number of blocks on each coordinate axis respectively; N m Indicates the number of samples in the geological model. i Yes An element, representing the well control sequence of the injection well, ni represents the number of injection wells, N c Indicates the number of control time steps; N w represents the number of well control sequences; w p yes An element of represents the well control sequence of the production well, np represents the number of production wells, u t Represents the start-up time step of the production well.

[0016] The specific method of establishing a U-FNO architecture suitable for CCS fast simulation and directly outputting pressure and saturation states in the time dimension is as follows:

[0017] The U-type network is connected in parallel with the Fourier neural operator network to obtain the U-FNO architecture, which is as follows: the input observation a(x) is promoted to a higher-dimensional feature space through a fully connected neural network, expressed as Next, a series of iterative FNO and U-FNO layers are used, which are able to capture the complex spatial dependencies in the data. These layers are used to train the function sequence and Finally, the architecture uses another fully connected neural network to project the high-dimensional output back to the original space, represented as

[0018] In each U-FNO layer, an integral kernel transformation K with learnable parameters, a U-Net CNN operator U, and a linear operator W are applied. The transformation is followed by a nonlinear activation function σ, resulting in a highly expressive model capable of approximating multiphase flows with less data requirements.

[0019] Among them, the Transformer module with self-attention mechanism is used to refine the output pressure and saturation states, and generate time-varying production sequences by capturing temporal dependencies.

[0020] Among them, the specific working process of the U-FNO architecture after introducing the Transformer module with self-attention mechanism is as follows:

[0021] After acquiring the pressure and saturation state images, convolutional layers are used for feature extraction to convert the state images into low-dimensional latent variables; these latent variables are then processed through the Transformer encoder to reconstruct the production rate sequence; the Transformer encoder component has a multi-layer architecture consisting of N f It consists of 10000 Transformer encoder layers, each of which is good at processing and refining the hidden sequence features of the input data;

[0022] The Transformer encoder layer mechanism is based on residual connections, supplemented by normalization layers, followed by multi-head self-attention layers, and finally multi-layer perceptrons; the multi-head self-attention layer is particularly effective in capturing complex dependencies within the sequence by paying attention to different positions in the input sequence in parallel; for a given sequence feature Z = (z1, z2, ..., z n ), the model uses linear layers to derive the corresponding transformation matrices Q, K, and V.

[0023] Among them, the overall performance analysis method is proposed, the error measurement mechanism is introduced, and the specific way to complete the simulation is as follows:

[0024] The simulation can be completed by proposing an overall performance analysis method, introducing an error measurement mechanism, and analyzing the training loss of the alternative model and the corresponding validation set root mean square error.

[0025] The invention discloses a CCS fast simulation method based on a multi-feature fusion neural network. The method constructs a U-FNO network which is a fusion of a U neural network (U-Net) and a Fourier neural operator (FNO), so as to improve the prediction accuracy of the pressure field and the saturation field in the long time series during the CO2 geological storage process. At the same time, a production series prediction network based on transformer encoders (TE) is adopted to directly connect the output end of the U-FNO network to the input end of the TE network, so as to realize the direct prediction from the geological parameters and the well control curve to the time series distribution of the pressure and saturation fields and the corresponding production curve. The method effectively overcomes the problem that the existing CCS simulation technology is difficult to process multi-feature input and output at the same time, improves the prediction efficiency of the influence of the CO2 storage rate on the reservoir safety, and can provide an efficient and reliable intelligent simulation tool for CCS engineering optimization and history matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 It is the overall framework of the agent modeling of the present invention.

[0028] Figure 2 3D configuration of the reservoir and well arrangement diagram of the embodiment of the present invention.

[0029] Figure 3 1 and 2 are six sample permeability graphs of a saline water reservoir according to an embodiment of the present invention.

[0030] Figure 4 These are six well control sequence samples according to an embodiment of the present invention.

[0031] Figure 5 is the training loss (MSE) and the corresponding RMSE on the validation set of the embodiment of the present invention.

[0032] Figure 6 All test samples of the embodiments of the present invention were evaluated by mean percentage absolute error (MPAE) and structural similarity index (SSIM).

[0033] Figure 7 3 is a diagram of pressure and gas saturation of the predicted reservoir (third layer) at the 6th and 16th control time steps according to an embodiment of the present invention.

[0034] Figure 8It is the pressure and gas saturation diagram prediction of the reservoir (sixth layer) at the 3rd, 6th and 16th control time steps of the embodiment of the present invention.

[0035] Fig. 9 It is a comparison between the simulation calculation results of the cumulative CO2 injection amount and the cumulative CO2 production amount of the embodiment of the present invention and the prediction results of the alternative model.

[0036] Fig.10 It is a cross plot of the predicted cumulative CO2 production and the actual value in all test samples of the embodiment of the present invention.

[0037] Fig.11 is the cumulative CO2 injection amount and the cumulative CO2 production amount of the embodiment of the present invention. 2 Box plot.

[0038] Fig.12 It is a flowchart of the steps of the CCS fast simulation method based on multi-feature fusion neural network of the present invention.

[0039] Fig.13 It is a basic reservoir parameter schematic diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0040] Embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0041] See also Figures 1 to 13 ,in Figure 1 is the overall framework of the agent modeling of the present invention, Figure 2 is a three-dimensional configuration of the reservoir and a well arrangement diagram of an embodiment of the present invention, Figure 3 are six permeability sample graphs of saline water reservoirs according to an embodiment of the present invention, Figure 4 are six well control sequence samples of the embodiment of the present invention, Figure 5 is the training loss (MSE) and the corresponding RMSE on the validation set of the embodiment of the present invention, Figure 6 All test samples of the embodiments of the present invention were subjected to MPAE and SSIM evaluation. Figure 7 is a diagram of pressure and gas saturation of the predicted cap layer (third layer) at the 3rd, 6th and 16th control time steps according to an embodiment of the present invention, Figure 8 is the pressure and gas saturation diagram prediction of the reservoir (sixth layer) at the 3rd, 6th and 16th control time steps of the embodiment of the present invention, Fig. 9 is a comparison between the simulation calculation results of the cumulative CO2 injection amount and the cumulative CO2 production amount of the embodiment of the present invention and the prediction results of the alternative model, Fig.10is a cross plot of the predicted cumulative CO2 production and the true value in all test samples of the embodiment of the present invention, Fig.11 is the cumulative CO2 injection amount and the cumulative CO2 production amount of the embodiment of the present invention. 2 Box plot, Fig.12 It is a flowchart of the steps of the CCS fast simulation method based on multi-feature fusion neural network of the present invention, Fig.13 It is a basic reservoir parameter schematic diagram of an embodiment of the present invention.

[0042] The present invention provides a CCS fast simulation method based on a multi-feature fusion neural network, comprising the following steps:

[0043] S100. Based on the determined simulation process, a U-FNO architecture suitable for CCS fast simulation is established, which can directly output pressure and saturation status in the time dimension;

[0044] For this specific implementation, the specific method of determining the simulation process is:

[0045] In the prediction of CO2 reservoir storage process under geological uncertainty, the well rate of injection and production wells is adjusted in each reservoir model at each control time step; the simulation process can be expressed as:

[0046] X=[P,S,q i ,q p ]=g(m,w i ,w p )

[0047] The geological model is represented by m, which is An element of N, where nb is defined as nx×ny×nz, which represents the total number of grid blocks in the geological model, and nx, ny, and nz correspond to the number of blocks on each coordinate axis respectively; N m Indicates the number of samples in the geological model. i Yes An element, representing the well control sequence of the injection well, ni represents the number of injection wells, N c Indicates the number of control time steps; N w represents the number of well control sequences; w p yes An element of represents the well control sequence of the production well, np represents the number of production wells, u t Represents the start-up time step of the production well.

[0048] The U-FNO architecture is as follows Figure 1 As shown in ①. The U-FNO approximation of the mapping includes the following steps:

[0049] The input observation a(x) is lifted to a higher-dimensional feature space through a fully connected neural network, expressed as Next, a series of iterative FNO and U-FNO layers are used, which are able to capture the complex spatial dependencies in the data. These layers are used to train the function sequence and Finally, the architecture uses another fully connected neural network to project the high-dimensional output back to the original space, represented as In each U-FNO layer, an integral kernel transformation K with learnable parameters, a U-Net CNN operator U, and a linear operator W are applied. The transformation is followed by a nonlinear activation function σ, resulting in a highly expressive model capable of approximating multiphase flows with less data requirements.

[0050] S200, based on the U-FNO architecture, introduces the Transformer module with self-attention mechanism;

[0051] For this specific implementation, a Transformer module with a self-attention mechanism is added, which refines these states and generates a time-varying production sequence by capturing temporal dependencies;

[0052] In this implementation, a Transformer module with a self-attention mechanism is added, and its architecture can be seen Figure 1 ②, the specific process is as follows:

[0053] After acquiring the pressure and saturation state images, convolutional layers are used for feature extraction to convert the state images into low-dimensional latent variables. These latent variables are then processed through the Transformer encoder to reconstruct the production rate sequence. The Transformer encoder component has a multi-layer architecture consisting of N f The network consists of 100 Transformer encoder layers, each of which is good at processing and refining the hidden sequence features of the input data.

[0054] The Transformer encoder layer mechanism is based on residual connections, supplemented by normalization layers, followed by multi-head self-attention (MHSA) layers, and finally multi-layer perceptrons (MLP). MHSA layers are particularly effective in capturing complex dependencies within the sequence by paying attention to different positions in the input sequence in parallel. For a given sequence feature Z = (z1, z2, ..., z n ), the model uses linear layers to derive the corresponding transformation matrices Q, K, and V. This approach is crucial for effectively processing sequence features and significantly improves the overall performance of the model.

[0055] S300, permeability and well control sequence were selected as model inputs, 10 permeability samples were randomly generated using SGeMS for sequential Gaussian simulation, and 500 groups of well control sequences were randomly generated for model training;

[0056] For this specific embodiment, in the present invention, SGeMS (Stanford Geostatistical Modeling Software) was used to perform sequential Gaussian simulation to randomly generate 10 permeability realizations and 500 groups of well control sequences were randomly generated.

[0057] S400. Finally, an overall performance analysis method is proposed, an error measurement mechanism is introduced, and the simulation is completed.

[0058] For this specific implementation, an overall performance analysis method is proposed, an error measurement mechanism is introduced, and the training loss of the alternative model and the corresponding validation set root mean square error (RMSE) are analyzed.

[0059] Take the simulation and analysis of CO2 storage for a three-dimensional (3D) actual saline aquifer reservoir as an example:

[0060] The model used in this example represents a three-dimensional (3D) actual saline aquifer reservoir, providing a realistic case for the simulation and analysis of CO2 storage. The reservoir domain is discretized into a Cartesian grid, divided into 101×135×6 grid cells along the x, y, and z axes, respectively, and each grid cell has a size of 40 meters by 40 meters in the horizontal plane. The top two layers of the model are caprocks, providing closure for the CO2 storage process, while the lower four layers are designated as saline aquifer reservoirs.

[0061] The reservoir model contains a total of 81,810 grid cells and has four injection wells and one production well, each of which vertically penetrates four layers below. The three-dimensional configuration of the reservoir and the well layout are shown in Figure 1. Figure 2 As shown, the basic reservoir parameters are detailed in Table 1.

[0062] Permeability samples were prepared for the surrogate model considering permeability and well control sequence as model inputs. In this example, 10 permeability realizations were randomly generated using sequential Gaussian simulation using SGeMS (Stanford Geostatistical Modeling Software). Figure 3 This reflects the uncertainty of the geological model.

[0063] In addition, this embodiment randomly generates 500 sets of well control sequences. Specifically, the well control sampling range of the injection well is set at 200,000m 3 / day up to 1,000,000m 3 / day. For production wells, activation times were sampled between 0 and 16 control time steps, and well control was sampled between 40MPa and 45MPa. Subsequently, for each model realization, 500 simulations were performed using the generated 500 sets of well control sequences, resulting in a total of 5,000 sets of samples. These constituted the training dataset. Figure 4 Six representative samples of well control sequences are shown. Pressure and saturation plots, as well as cumulative CO2 injection and CO2 production from production wells, are calculated using the reservoir numerical simulation CMG-GEM. Simulation results for all samples are generated and saved at each time step. The simulation time step is set to 180 days, and the total simulation time is 180 days multiplied by 16. Well control is adjusted every 180 days, with a total of 16 control steps during the entire simulation. These datasets are then used to build and train proxy models for each layer of the CO2 saline reservoir, resulting in 14 proxy models. The 14 models are trained in parallel, each model undergoing 50 epochs, with a learning rate of 0.001 and a batch size of 8. Training is performed on eight NVIDIA TeslaA800 GPUs, each with 80GB of memory. The entire training process took a total of 625.5 seconds.

[0064] Figure 5 The training loss and its root mean square validation error (RMSE) of the surrogate model at each layer are shown. Layers 1 and 2 act as cap layers, while layers 3 to 6 constitute the reservoir. It is worth noting that layer 2, as part of the cap layer, is connected to layer 3 of the reservoir. The significant difference in geological parameters between these two layers leads to higher training loss and RMSE, indicating that its training performance and generalization ability are poorer than other layers. This suggests that the abrupt change in geological properties at the interface between the cap layer and the reservoir may pose a challenge to the effective learning of the model.

[0065] The present invention also evaluates the generalization performance of the model on the test set. Figure 6 The mean percentage absolute error (MAPE) and structural similarity index (SSIM) evaluation results for the entire test set are shown, and the gray area forms all samples. The second, fourth, and sixth layers of the test case are selected for display. In addition, Fig.11 R 2 In the box plots, for all samples, the SSIM values ​​of pressure and saturation are greater than 0.95 and 0.987, respectively, and the MAPE values ​​are less than 0.04 and 0.025, respectively. These evaluation results show that the error of most samples remains within 0.06, proving that the reconstruction of image features is suitable for practical CO2 storage prediction applications.

[0066] The present invention also expands the scope of investigation to include the ability to predict pressure and saturation distribution in 3D saline reservoirs after CO2 storage. To evaluate the model performance, the simulation results were compared with the model's predictions under various random well control sequences and random permeability realizations. Figure 7 The pressure and saturation fields of the caprock reservoir (third layer) after CO2 storage are shown, indicating that the model can accurately depict the state of the migrating plume from the CO2 reservoir. Figure 8 The state of CO2 plume in the reservoir (sixth layer) is shown, and the colored bar indicates the absolute error of the entire reservoir pressure and saturation prediction, which are controlled within 0.05MPa and 0.035 respectively.

[0067] Near the CO2 displacement front, the prediction error of saturation is larger than the error inside the CO2 injection area, and the error distribution near the front is also affected by heterogeneity. The results clearly show that the CO2 saturation prediction error of the model is mainly concentrated at the CO2 displacement front. This error distribution is a direct result of the training process using the MSE loss function. The MSE loss function amplifies the penalty of errors in areas with higher saturation and preferentially reduces the error in areas with higher saturation inside the CO2 injection area.

[0068] Fig. 9 The CO2 injection and production volumes calculated by the surrogate model within 16 control time steps are shown. The dashed line in each sub-figure represents the calculated output of the full physical simulation, while the dot-dash line represents the prediction result of the surrogate model. It can be seen that the CO2 injection and production volumes of the wells simulated by the surrogate model are in good agreement with the results of the full physical simulation. However, it should be noted that the prediction results of the surrogate model show certain irregularities compared with the smooth curves generated by the full physical process. The prediction differences between the surrogate model and the full physical simulation arise from the inherent characteristics of the deep learning network, which introduces different degrees of error in the prediction of pressure and saturation, resulting in uneven results of the surrogate model in predicting CO2 injection and production volumes.

[0069] The present invention shows a cross-plot of the cumulative CO2 injection and production volumes for the surrogate model and the full physical simulation. Fig.10 Comparison results for all test cases at the 3rd, 6th, and 16th control time steps are shown. The surrogate model is generally consistent with the calculated output of the full physical model in terms of cumulative CO2 injection and CO2 production rate in production wells. However, the difference between the surrogate model and the full physical model becomes more pronounced as the cumulative CO2 injection rate increases. The increase in the number of active production wells in the test sample leads to significant changes in the reservoir pressure and saturation response as CO2 injection progresses. This increase in complexity makes it more difficult for the surrogate model to learn accurately, ultimately resulting in the predicted differences in cumulative CO2 injection and production volume widening over time.

[0070] Fig.11 The coefficient of determination (R 2 ). In each box plot, the center line represents the median of the coefficient of determination, the bottom and top of the box correspond to the 25th and 75th percentiles, respectively, and the whiskers represent the minimum and maximum values ​​observed. Median R of cumulative CO2 injected and CO2 produced volumes 2 are 0.972 and 0.98, respectively. Overall, the U-FNO+Transformer network can efficiently and accurately predict the pressure and saturation states, and can accurately predict the cumulative CO2 injection and production volumes throughout the life cycle of the saline aquifer reservoir.

[0071] Experiments show that this method can effectively predict production dynamics in studying CCS problems.

[0072] A CCS fast simulation method based on a multi-feature fusion neural network of the present invention is used. When it is used specifically, the method constructs a U-FNO network that integrates a U neural network (U-Net) and a Fourier neural operator (FNO) to improve the prediction accuracy of the pressure field and the saturation field in the long time series during the CO2 geological storage process; at the same time, a production series prediction network based on transformer encoders (TE) is adopted to directly connect the output end of the U-FNO network to the input end of the TE network to achieve direct prediction from geological parameters and well control curves to the time series distribution of pressure and saturation fields and the corresponding production curves; this method effectively overcomes the problem that the existing CCS simulation technology is difficult to simultaneously process multi-feature inputs and outputs, improves the prediction efficiency of the impact of CO2 storage rate on reservoir safety, and can provide an efficient and reliable intelligent simulation tool for CCS engineering optimization and history matching.

[0073] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A CCS fast simulation method based on multi-feature fusion neural network, characterized in that: The steps include: Establish a U-FNO architecture suitable for CCS fast simulation and direct output of pressure and saturation status in the time dimension; Based on the U-FNO architecture, a Transformer module with self-attention mechanism is introduced to obtain a fast simulation alternative model for building CCS; Permeability and well control sequences were selected as model inputs, and SGeMS was used to perform sequential Gaussian simulation to randomly generate 10 permeability samples and 500 sets of well control sequences for model training; Finally, an overall performance analysis method is proposed, an error measurement mechanism is introduced, and the simulation is completed.

2. The CCS fast simulation method based on multi-feature fusion neural network according to claim 1 is characterized in that: When building a surrogate model: In the prediction of CO2 reservoir storage process under geological uncertainty, the well control of injection and production wells is adjusted in each reservoir model at each control time step; the mapping relationship required to be achieved by the relevant alternative model is expressed as: X=[P,S,q i ,q p ]=g(m,w i ,w p ) The geological model is represented by m, which is An element of N, where nb is defined as nx×ny×nz, which represents the total number of grid blocks in the geological model, and nx, ny, and nz correspond to the number of blocks on each coordinate axis respectively; N m represents the number of samples in the geological model; w i Yes An element, representing the well control sequence of the injection well, ni represents the number of injection wells, N c Indicates the number of control time steps; N w represents the number of well control sequences; w p yes An element of represents the well control sequence of the production well, np represents the number of production wells, u t Represents the start-up time step of the production well.

3. The CCS fast simulation method based on multi-feature fusion neural network as claimed in claim 2 is characterized in that: The specific method to establish a U-FNO architecture suitable for CCS fast simulation and directly output pressure and saturation status in the time dimension is: The U-type network is connected in parallel with the Fourier neural operator network to obtain the U-FNO architecture, which is as follows: the input observation a(x) is promoted to a higher-dimensional feature space through a fully connected neural network, expressed as Next, a series of iterative FNO and U-FNO layers are used, which are able to capture the complex spatial dependencies in the data; these layers are used to train the function sequence and Transform, operating on the channel dimension c; finally, the architecture uses another fully connected neural network to project the high-dimensional output back to the original space, represented as In each U-FNO layer, an integral kernel transformation K with learnable parameters, a U-Net CNN operator U, and a linear operator W are applied; the transformation is followed by a nonlinear activation function σ, resulting in a highly expressive model capable of approximating multiphase flows with less data requirements.

4. The CCS fast simulation method based on multi-feature fusion neural network as claimed in claim 3 is characterized in that: Transformer modules with self-attention mechanism are used to refine the output pressure and saturation states and generate time-varying production sequences by capturing temporal dependencies.

5. The CCS fast simulation method based on multi-feature fusion neural network as claimed in claim 4 is characterized in that: The specific working process of the U-FNO architecture after introducing the Transformer module with self-attention mechanism is as follows: After acquiring the pressure and saturation state images, convolutional layers are used for feature extraction to convert the state images into low-dimensional latent variables; these latent variables are then processed through the Transformer encoder to reconstruct the production rate sequence; the Transformer encoder component has a multi-layer architecture consisting of N f It consists of 10000 Transformer encoder layers, each of which is good at processing and refining the hidden sequence features of the input data; The Transformer encoder layer mechanism is based on residual connections, supplemented by normalization layers, followed by multi-head self-attention layers, and finally multi-layer perceptrons; the multi-head self-attention layer is particularly effective in capturing complex dependencies within the sequence by paying attention to different positions in the input sequence in parallel; for a given sequence feature Z = (z1, z2, ..., z n ), the model uses linear layers to derive the corresponding transformation matrices Q, K, and V.

6. The CCS fast simulation method based on multi-feature fusion neural network as claimed in claim 5 is characterized in that: The overall performance analysis method is proposed, and the error measurement mechanism is introduced. The specific way to complete the simulation is as follows: The simulation can be completed by proposing an overall performance analysis method, introducing an error measurement mechanism, and analyzing the training loss of the alternative model and the corresponding validation set root mean square error.

Images