AI optimization-based carbon capture solvent regeneration energy consumption real-time regulation and control system and method

Abstract

The invention relates to the technical field of carbon capture regulation and control, and discloses a carbon capture solvent regeneration energy consumption real-time regulation and control system and method based on AI optimization. The system comprises a solvent regeneration environment monitoring unit which is used for collecting multi-dimensional operation parameters such as reboiler steam pressure, solvent circulation flow, lean and rich liquid heat exchange temperature difference and tower top gas phase component concentration in a regeneration tower in real time; the dynamic energy consumption feature extraction unit is used for extracting an implicit energy consumption feature vector through a deep space-time convolutional neural network and generating an energy efficiency evaluation index; the regeneration strategy optimization decision-making unit is used for generating a steam valve opening degree adjusting strategy and a solvent circulating pump frequency control strategy through online iteration by adopting a reinforcement learning framework; the execution feedback control unit is used for converting the strategy into a control signal and acquiring execution feedback data; and the closed-loop verification module is used for comparing the actual steam consumption deviation with the theoretical steam consumption deviation, and triggering feature re-extraction when the deviation exceeds a threshold value. The system realizes intelligent real-time regulation and control of solvent regeneration energy consumption.

Description

Technical Field

[0001] This invention relates to the field of carbon capture and regulation technology, specifically to an AI-optimized real-time control system and method for carbon capture solvent regeneration energy consumption. Background Technology

[0002] In the process of industrial carbon emission reduction, carbon capture technology is an important means of reducing carbon emissions, among which solvent absorption is widely used due to its wide applicability. Solvent regeneration, as a key step in this method, directly affects the economics and feasibility of the carbon capture process due to its energy consumption level. Currently, solvent regeneration mainly relies on the heating process within the regeneration tower to achieve solvent recycling, while reboiler steam consumption is the main source of energy consumption.

[0003] Traditional solvent regeneration control often employs experience-based, fixed-parameter control methods, which struggle to cope with energy consumption fluctuations caused by changes in operating conditions. For example, when the composition, flow rate, or solvent concentration of the feed gas fluctuates, the fixed steam valve opening and solvent circulation flow rate cannot match real-time demands, easily leading to excessive steam consumption or incomplete regeneration. Furthermore, existing control systems lack the ability to comprehensively analyze multi-dimensional operating parameters, relying solely on single-parameter feedback for adjustment. This makes it difficult to capture the coupling relationships between parameters, limiting the accuracy of energy consumption control.

[0004] Conventional energy consumption assessments often rely on offline data analysis, failing to reflect the energy efficiency status of the regeneration process in real time, leading to a lag in adjustments to control strategies. In actual operation, parameters such as the temperature difference between the lean and rich liquids in the regeneration tower and the concentration of gaseous components at the top of the tower dynamically change. Traditional control systems struggle to respond quickly to these changes and generate optimized strategies, resulting in widespread energy waste. As carbon capture scales expand, the need for refined control of solvent regeneration energy consumption becomes increasingly urgent, and existing control methods are no longer sufficient to meet the requirements of efficient and low-carbon operation. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-optimized real-time control system for carbon capture solvent regeneration energy consumption, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an AI-optimized real-time energy consumption control system for carbon capture solvent regeneration, the system comprising:

[0007] The solvent regeneration environment monitoring unit is used to collect multi-dimensional operating parameters in the carbon capture solvent regeneration tower in real time. The multi-dimensional operating parameters include at least the reboiler steam pressure data, solvent circulation flow rate data, lean and rich liquid heat exchange temperature difference data, and tower top gas phase component concentration data.

[0008] The dynamic energy consumption feature extraction unit is used to receive the multi-dimensional operation parameters and extract the implicit energy consumption feature vector of the solvent regeneration process through a deep spatiotemporal convolutional neural network, while generating a solvent regeneration energy efficiency evaluation index.

[0009] The regeneration strategy optimization decision unit is used to receive the implicit energy consumption feature vector and the solvent regeneration energy efficiency evaluation index, and to use a reinforcement learning framework to iteratively generate a steam valve opening adjustment strategy and a solvent circulation pump frequency control strategy online.

[0010] The execution feedback control unit is used to convert the steam valve opening adjustment strategy into a pulse control signal for the steam regulating valve, and at the same time convert the solvent circulation pump frequency control strategy into a speed control signal for the variable frequency drive, and to acquire valve displacement feedback data and pump speed feedback data in real time.

[0011] The closed-loop verification module is used to compare the deviation between the actual steam consumption data and the theoretical steam consumption data during the verification period. When the deviation exceeds the preset tolerance threshold, the dynamic energy consumption feature extraction unit is triggered to re-extract the implicit energy consumption feature vector of the solvent regeneration process.

[0012] Preferably, the dynamic energy consumption feature extraction unit specifically includes:

[0013] The parallel convolution processing module receives the reboiler steam pressure time series data and solvent circulation flow time series data transmitted by the solvent regeneration environment monitoring unit, and extracts the steam pressure fluctuation feature map and flow fluctuation feature map respectively through a dual-channel one-dimensional convolution layer.

[0014] The spatiotemporal correlation analysis module receives the steam pressure fluctuation feature map and flow fluctuation feature map output by the parallel convolution processing module, and uses a three-dimensional convolution kernel to fuse the spatiotemporal dimension correlation features to output a steam-flow coupling feature tensor.

[0015] The energy efficiency assessment generation module receives the steam-flow coupling characteristic tensor transmitted by the spatiotemporal correlation analysis module, combines the lean and rich liquid heat exchange temperature difference data with the column top gas phase component concentration data, and generates solvent regeneration energy efficiency assessment indicators through full-connected layer regression.

[0016] Preferably, the regeneration strategy optimization decision unit performs the following operations:

[0017] Establish a strategy decision-making time window and obtain the solvent regeneration energy efficiency evaluation index output by the dynamic energy consumption feature extraction unit at a fixed sampling period;

[0018] The solvent regeneration energy efficiency evaluation index in the current time window and the solvent regeneration energy efficiency evaluation index in the historical time window are input into the long short-term memory network to predict the optimal energy consumption reduction gradient for the next three operating cycles.

[0019] Based on the predicted optimal energy consumption descent gradient, a dual-action output strategy network is used to generate the percentage change of the steam valve opening adjustment strategy and the frequency adjustment of the solvent circulation pump frequency control strategy.

[0020] Preferably, the execution feedback control unit includes:

[0021] The valve signal conversion module receives the percentage change of the steam valve opening adjustment strategy transmitted by the regeneration strategy optimization decision unit, and generates a steam regulating valve pulse control signal with a corresponding duty cycle through the pulse width modulation circuit.

[0022] The variable frequency drive module receives the frequency adjustment amount of the solvent circulation pump frequency control strategy transmitted by the regeneration strategy optimization decision unit, and generates a stepped speed control signal for the variable frequency drive through a digital signal processor.

[0023] The displacement sensing and acquisition module monitors the valve stem displacement of the steam regulating valve in real time and converts it into valve execution displacement feedback data.

[0024] The speed feedback acquisition module captures the real-time speed value of the solvent circulation pump motor through a Hall sensor and converts it into pump speed feedback data.

[0025] Preferably, the operation of the closed-loop verification module includes:

[0026] Obtain historical data on the opening degree of the steam regulating valve and the operating frequency of the solvent circulation pump recorded by the execution feedback control unit;

[0027] The system calls upon the implicit energy consumption feature vectors for the corresponding time period stored in the dynamic energy consumption feature extraction unit, and calculates the theoretical steam consumption data through the forward propagation network.

[0028] The absolute deviation between the actual steam consumption data and the theoretical steam consumption data is compared. When the absolute deviation exceeds the preset tolerance threshold for three sampling cycles, a feature re-extraction instruction is sent to the dynamic energy consumption feature extraction unit.

[0029] Preferably, the system further includes:

[0030] The solvent state analysis unit is connected between the solvent regeneration environment monitoring unit and the dynamic energy consumption feature extraction unit. It is used to receive solvent circulation flow data and column top gas phase component concentration data, and output solvent chemical stability index through solvent degradation degree analysis model.

[0031] The dynamic energy consumption feature extraction unit uses solvent chemical stability index as an additional feature dimension, and inputs it into the deep spatiotemporal convolutional neural network along with multi-dimensional operating parameters.

[0032] Preferably, the solvent state analysis unit includes:

[0033] The solvent composition analysis module receives the concentration data of gas phase components at the top of the column and analyzes the concentration distribution of solvent degradation products through a gas chromatography simulation algorithm.

[0034] The stability assessment module receives data on the concentration distribution of solvent degradation products and solvent circulation flow rate, and generates solvent chemical stability indicators based on preset solvent chemical bond breaking thresholds.

[0035] Preferably, the system further includes:

[0036] A multi-channel heat exchange controller is connected between the regeneration strategy optimization decision unit and the execution feedback control unit. It is used to receive the steam valve opening adjustment strategy and the solvent circulation pump frequency control strategy, and calculate the flow distribution coefficient of each heat exchange branch through the thermodynamic balance model.

[0037] The execution feedback control unit corrects the amplitude of the steam regulating valve pulse control signal according to the flow distribution coefficient.

[0038] Preferably, the multi-channel heat exchange controller performs:

[0039] Receive the target steam flow value of the current steam valve opening regulation strategy transmitted by the regeneration strategy optimization decision unit;

[0040] The heat exchange temperature difference data of lean and rich liquids collected by the solvent regeneration environment monitoring unit is obtained, and the theoretical flow rate ratio of each heat exchange branch is calculated by the heat transfer coefficient compensation algorithm.

[0041] The flow distribution coefficient for each heat exchange branch is generated by multiplying the target steam flow rate by the theoretical flow rate ratio.

[0042] Preferably, the present invention further includes an AI-optimized method for real-time control of carbon capture solvent regeneration energy consumption, applied to the above-mentioned system, the method comprising:

[0043] The solvent regeneration environment monitoring unit captures real-time data on reboiler steam pressure, solvent circulation flow rate, lean and rich liquid heat exchange temperature difference, and column top vapor phase component concentration.

[0044] The captured multi-dimensional operation parameters are input into the dynamic energy consumption feature extraction unit, and the hidden energy consumption feature vector and solvent regeneration energy efficiency evaluation index are generated by using a deep spatiotemporal convolutional neural network.

[0045] A reinforcement learning framework is used to process the implicit energy consumption feature vector and the solvent regeneration energy efficiency evaluation index, and the steam valve opening adjustment strategy and solvent circulation pump frequency control strategy are output online iteratively.

[0046] The steam valve opening regulation strategy is converted into a steam regulating valve pulse control signal by executing the feedback control unit, and the solvent circulation pump frequency control strategy is converted into a variable frequency drive speed control signal.

[0047] Real-time acquisition of valve displacement feedback data and pump speed feedback data, combined with theoretical steam consumption data, to perform a closed-loop verification process;

[0048] When the deviation between the actual steam consumption data and the theoretical steam consumption data exceeds the preset tolerance threshold, the dynamic energy consumption feature extraction unit is triggered to re-extract the implicit energy consumption feature vector.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] This AI-optimized real-time energy consumption control system for carbon capture solvent regeneration achieves intelligent control of the solvent regeneration process through multi-unit collaborative operation. The solvent regeneration environmental monitoring unit can collect multi-dimensional operating parameters within the regeneration tower in real time, including reboiler steam pressure, solvent circulation flow rate, lean-rich liquid heat exchange temperature difference, and tower top gas phase component concentration, providing comprehensive basic data for subsequent energy consumption analysis and strategy optimization. This real-time data accurately reflects the dynamic operating conditions within the regeneration tower, allowing the system to promptly detect subtle changes in various parameters.

[0051] The dynamic energy consumption feature extraction unit utilizes a deep spatiotemporal convolutional neural network to process multi-dimensional operational parameters, extracting implicit energy consumption feature vectors and generating energy efficiency evaluation indicators. This approach can deeply uncover the potential correlations between parameters, capturing energy consumption variation patterns that are difficult to identify using traditional analysis methods, enabling the system to have a more accurate understanding of the energy consumption characteristics of the regeneration process. The generated energy efficiency evaluation indicators can intuitively reflect the current energy efficiency status of the regeneration process, providing an effective basis for subsequent strategy optimization.

[0052] The regeneration strategy optimization decision unit employs a reinforcement learning framework, combining implicit energy consumption feature vectors with energy efficiency evaluation indicators to iteratively generate steam valve opening adjustment strategies and solvent circulation pump frequency control strategies online. The application of reinforcement learning enables the system to continuously optimize its control strategies through constant interaction with the environment, adapting to energy consumption control requirements under different operating conditions. This online iterative approach allows the control strategies to closely follow the dynamic changes in the regeneration process, avoiding the limitations of traditional fixed strategies.

[0053] The execution feedback control unit translates the optimization strategy into specific control signals and acquires feedback data from the actuators in real time, forming a closed-loop link from strategy to execution. Feedback data promptly reflects the execution effect, allowing the system to understand the actual effectiveness of the strategy and providing a reference for further strategy adjustments. The closed-loop verification module compares the deviation between actual and theoretical steam consumption. When the deviation exceeds a threshold, it triggers feature re-extraction, ensuring that the system's understanding of energy consumption characteristics remains consistent with the actual process. When significant changes in operating conditions lead to alterations in energy consumption characteristics, this module can promptly activate an update mechanism to maintain the system's regulatory adaptability and reduce energy consumption anomalies caused by feature mismatch.

[0054] The coordinated operation of each unit enables real-time sensing of multi-dimensional parameters, dynamic extraction of deep features, online generation of optimization strategies, and real-time verification of execution effects in the energy consumption control of the solvent regeneration process. This forms a complete intelligent control system that effectively addresses the energy consumption control challenges brought about by operating condition fluctuations and improves the energy consumption management level of the solvent regeneration process. Attached Figure Description

[0055] Figure 1 This is a schematic diagram illustrating the working principle of the AI-optimized real-time energy consumption control system for carbon capture solvent regeneration described in this invention.

[0056] Figure 2 This is a flowchart illustrating the working principle of the dynamic energy consumption feature extraction unit.

[0057] Figure 3 Flowchart illustrating the working principle of the decision-making unit for regeneration strategy optimization;

[0058] Figure 4 This is a flowchart illustrating the working principle of the carbon capture solvent regeneration energy consumption control system for a solvent-containing state analysis unit. Detailed Implementation

[0059] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Please see Figure 1 This invention provides an AI-optimized real-time energy consumption control system and method for carbon capture solvent regeneration, the system comprising:

[0061] The system comprises a solvent regeneration environment monitoring unit, a dynamic energy consumption feature extraction unit, a regeneration strategy optimization decision-making unit, an execution feedback control unit, and a closed-loop verification module. The solvent regeneration environment monitoring unit collects real-time operational parameters such as reboiler steam pressure, solvent circulation flow rate, lean-rich liquid heat exchange temperature difference, and top gas phase component concentration using pressure sensors, flow meters, temperature probes, and gas analyzers distributed within the regeneration tower. The dynamic energy consumption feature extraction unit processes these parameters using a deep spatiotemporal convolutional neural network to extract implicit energy consumption feature vectors containing steam-solvent interaction characteristics and outputs energy efficiency evaluation indicators. The regeneration strategy optimization decision-making unit, based on a reinforcement learning framework, dynamically generates steam valve opening adjustment strategies and pump frequency control strategies according to the feature vectors and energy efficiency indicators. The execution feedback control unit converts the strategies into executable control signals to drive field equipment through a pulse width modulation circuit and a digital signal processor. The closed-loop verification module continuously compares actual and theoretical steam consumption; when the deviation exceeds a threshold, a feature re-extraction mechanism is triggered, forming an adaptive optimization closed loop.

[0062] Example 1: See Figure 2 The dynamic energy consumption feature extraction unit, as the core computing module of the system, employs a deep spatiotemporal convolutional neural network architecture to process multi-dimensional operational parameters from the solvent regeneration environment monitoring unit. This unit achieves dynamic extraction and energy efficiency assessment of energy consumption characteristics during the solvent regeneration process through parallel processing and feature fusion mechanisms.

[0063] At the hardware implementation level, this unit is deployed on an industrial edge computing device, equipped with an NVIDIA Jetson AGX Orin processor, featuring 256 CUDA cores and 64GB of video memory. Data processing employs dual-buffering technology, using two independent DDR4 memory areas to alternately receive monitoring data, ensuring continuous processing of real-time data streams. The system clock maintains millisecond-level synchronization with the field PLC, achieving timing alignment between data acquisition and control commands through the IEEE 1588 precision time protocol.

[0064] The parallel convolutional processing module is designed with a dual-channel structure, processing the reboiler steam pressure time series and the solvent circulation flow time series separately. Each channel is configured with three sets of parallel convolutional kernels, corresponding to three time window widths of 5 seconds, 7 seconds, and 9 seconds. The convolutional layers adopt a causal convolution mode, using only current and historical data for computation to avoid future information leakage. Each convolutional group contains 32 filters with a stride of 1 second, using SAME padding to maintain the temporal length of the feature map. After the convolution operation, a max pooling layer is applied, with a pooling window width of 3 seconds and a stride of 2 seconds, to achieve feature dimensionality reduction. A batch normalization layer introduces learnable scaling and offset parameters to normalize each feature channel independently. The two channels ultimately output a 64-dimensional feature map, which is aligned to the time dimension with zero padding and then sent to the spatiotemporal correlation analysis module.

[0065] The spatiotemporal correlation analysis module employs a 3D convolutional structure to process spatiotemporal features. The input feature maps are first stacked along the channel dimension to form a 128×T×2 tensor (T being the time step). The 3D convolutional kernel size is set to 3×3×3, corresponding to the time, space, and channel dimensions. The module comprises four processing stages: the first stage uses 64 3D convolutional kernels to extract local spatiotemporal features; the second stage introduces a 3D attention mechanism to calculate channel-time joint attention weights; the third stage performs cross-channel feature fusion, achieving feature recombination through 1×1×1 convolutions; and the fourth stage uses 3D dilated convolutions to expand the receptive field, with a dilation rate set to 2. Each processing stage is followed by 3D batch normalization and the LeakyReLU activation function, ultimately outputting a 128-dimensional steam-flow coupled feature tensor.

[0066] The energy efficiency assessment generation module employs a feature concatenation and fully connected network structure. The heat exchange temperature difference data between lean and rich liquids, after being filtered by moving average, is combined with the concentration data of gas phase components at the top of the column to form an auxiliary feature vector. This vector is extended to 64 dimensions through a linear transformation layer and concatenated with the feature tensor output by the spatiotemporal correlation analysis module in the feature dimension, forming a 192-dimensional composite feature vector. The fully connected network contains three hidden layers with 256, 128, and 64 neurons respectively. Each layer uses Dropout regularization (dropout rate 0.2) to prevent overfitting. The output layer uses a Sigmoid activation function to compress the energy efficiency assessment index to the 0-1 range, corresponding to the thermodynamic perfection of the solvent regeneration process.

[0067] During real-time computation, the system establishes a dynamic feature caching mechanism. Intermediate features from each processing step are stored in HBM2 high-speed video memory, with data sharing between modules achieved through memory-mapped files. The computation graph employs a static optimization approach, utilizing TensorRT for graph optimization and layer fusion to keep computation latency below 50 milliseconds. Anomaly detection is implemented during feature extraction; when input data exceeds a preset reasonable range, the system automatically switches to backup data or activates prediction mode to maintain operation.

[0068] The unit incorporates a feature visualization interface, connected to an HMI (Human-Machine Interface) via an industrial switch. Steam pressure and flow rate fluctuation feature maps are displayed as heatmaps, while spatiotemporal correlation features are presented as 3D point clouds, enabling operators to intuitively understand the feature extraction process. All feature data is simultaneously written to a time-series database with a storage period of 30 days for offline analysis and model optimization.

[0069] Network parameter updates employ an online learning strategy. The system maintains two parallel model instances: the main model processes real-time data, while the shadow model receives the latest 30 days of historical data for incremental training. When the shadow model's loss function value on the validation set is superior to the main model, a model switching mechanism is automatically triggered. The training process uses stochastic gradient descent with momentum, a learning rate of 0.001, a momentum parameter of 0.9, and processes data from 256 time steps per batch.

[0070] The interface between the unit and subsequent modules uses the Protobuf protocol to encapsulate data. Implicit energy consumption feature vectors are transmitted in float32 array format, accompanied by timestamps and confidence scores. Energy efficiency assessment indicators include three sub-items: current value, moving average, and trend slope, each with a data quality flag. The communication link implements CRC checksum and retransmission mechanisms to ensure data transmission reliability.

[0071] In terms of power management, the unit is equipped with dual redundant power supply modules, supporting both 24VDC and 220VAC inputs. The computing core employs dynamic voltage and frequency adjustment technology, automatically adjusting the computing frequency according to the processing load to balance real-time performance and energy consumption. The cooling system uses a combination of heat pipes and centrifugal fans to ensure continuous and stable operation under industrial ambient temperatures.

[0072] Example 2: See Figure 3 The regeneration strategy optimization decision unit, based on a reinforcement learning framework, dynamically generates steam valve opening adjustment strategies and solvent circulation pump frequency control strategies to achieve optimal energy consumption control in the solvent regeneration process. This unit employs time series segmentation, long short-term memory network prediction, and a dual-action output strategy network to ensure the real-time performance and adaptability of the control strategy.

[0073] In terms of hardware architecture, this unit is deployed on an industrial control server, equipped with an Intel Xeon Scalable processor and an NVIDIA Tesla T4 accelerator card, supporting parallel computing and low-latency inference. Data processing employs a circular buffer structure to store the operation data of the most recent 30 minutes, ensuring that policy decisions are based on the latest system state. The system clock is synchronized with the DCS (Distributed Control System) and maintains microsecond-level time accuracy via the IEEE 1588 protocol, guaranteeing the timing consistency of control commands.

[0074] The strategy decision-making time window is set to 180 seconds, matching the scanning cycle of the on-site PLC. The data preprocessing module applies a moving average filter to the input time series to eliminate short-term noise interference while preserving long-term trend characteristics. The Long Short-Term Memory (LSTM) network employs a two-layer stacked structure, with each layer containing 128 neurons. The input sequence length is fixed at 6 historical windows (a total of 1080 seconds of data). Network training uses a teacher-forcing strategy, employing real historical data as input during the training phase and an autoregressive prediction mode during the inference phase. The output layer connects to a time-distributed fully connected layer, predicting the optimal energy consumption descent gradient curve for the next 540 seconds. The prediction results are output at 10-second intervals, forming a continuous control reference trajectory.

[0075] The dual-action output strategy network consists of two parallel subnetworks, responsible for regulating the steam valve opening and the solvent circulation pump frequency, respectively. The valve control subnetwork employs a dual-branch architecture: the main branch processes the current implicit energy consumption feature vector and energy efficiency evaluation indicators, while the auxiliary branch receives feedback on the execution effects of historical control commands (including valve displacement deviation and steam consumption deviation). The features from the two branches are fused in the intermediate layer and output as the valve opening adjustment amount after passing through three fully connected layers (with 256, 128, and 64 neurons respectively). The pump frequency control subnetwork introduces a residual connection structure, where the output of each layer is added to the input before being passed to the next layer, avoiding the gradient vanishing problem. Finally, the network uses a Tanh activation function to limit the output to the range [-1, 1], and then maps it to the actual adjustment amount (valve opening variation range ±15%, pump frequency adjustment range ±5Hz).

[0076] Policy optimization employs the Proximal Policy Optimization (PPO) algorithm, performing online updates every 60 seconds. The algorithm maintains an experience replay buffer, storing the state-action-reward triplet for the most recent 24 hours. During training, the policy network and environment model optimize alternately: the policy network generates control commands, while the environment model simulates the system response and calculates the reward value. The reward function comprehensively considers energy consumption descent gradients, operational stability (the magnitude of changes in valve and pump frequencies), and system constraints (such as pressure limits within the tower). Training uses the Adam optimizer with a learning rate of 0.0001 and a policy update step size limited to within 0.01 to ensure smooth policy adjustment.

[0077] The control strategy output adopts an incremental adjustment mode. Within each decision cycle, the strategy network generates a relative adjustment amount, rather than an absolute setpoint, to avoid large fluctuations. The steam valve opening adjustment amount is output as a percentage change, which is then limited before being sent to the actuator. The solvent circulation pump frequency control strategy adopts a gradual adjustment, fine-tuning the frequency every 10 seconds to avoid impacting the motor drive system. All control commands are accompanied by timestamps and priority identifiers to ensure that reasonable control logic is maintained even in the event of communication delays or data loss.

[0078] An internal strategy evaluation module is incorporated into the unit to monitor control effectiveness in real time. This module calculates the deviation between the actual energy consumption reduction rate and the predicted value; when the deviation exceeds a threshold, a strategy recalculation is triggered. Simultaneously, the module records the adjustment frequency of valves and pumps to prevent wear and tear on actuators due to frequent operation. Historical strategy data is stored in a time-series database, supporting offline analysis and strategy optimization.

[0079] Regarding the communication interface, the unit uses the OPCUA protocol to interact with the execution feedback control unit. Control commands are transmitted in a structured data format, including target values, adjustment ranges, and expected execution times. Feedback data (such as actual valve displacement and pump speed) is transmitted back through the same channel, forming a closed-loop control. The communication link implements data encryption and integrity verification to prevent unauthorized access or tampering.

[0080] The power supply and heat dissipation design takes into account the stringent requirements of industrial environments. The unit employs redundant power modules, supporting dual inputs of 24VDC and 220VAC, and switches to UPS power in the event of a power outage. The compute nodes are equipped with a liquid cooling system to ensure continuous and stable operation in high-temperature environments. All electronic components comply with industrial-grade EMC standards, resisting electromagnetic interference in the field.

[0081] The system supports both online learning and manual intervention modes. In the default mode, the policy network makes decisions entirely autonomously; under special circumstances, operators can adjust policy weights or directly input control commands through the HMI interface. Manual intervention records are marked and stored for subsequent model training and data playback analysis.

[0082] The implementation of this unit enables adaptive energy control of the solvent regeneration process, allowing for dynamic strategy adjustments based on real-time operating conditions while maintaining operational stability. Through continuous optimization of the reinforcement learning framework, the system gradually improves control accuracy and energy efficiency over long-term operation.

[0083] Example 3: The execution feedback control unit, as a key component of system command execution and status feedback, is responsible for converting optimization strategies into actual control signals and acquiring equipment operating status. This unit adopts a multi-channel signal conversion architecture to achieve high-precision execution and real-time feedback of control commands, forming a complete control closed loop.

[0084] In terms of hardware implementation, this unit adopts a modular design, including a valve signal conversion module, a frequency converter drive module, a displacement sensing and acquisition module, and a speed feedback acquisition module. The modules are interconnected via a backplane bus, and the communication protocol adopts the CANopen industrial standard. The main controller uses an STM32H743 microprocessor, equipped with a hardware floating-point unit to ensure real-time signal conversion. The power supply system adopts an isolated DC-DC conversion scheme, with the input 24VDC filtered and regulated before being distributed to each functional module.

[0085] The valve signal conversion module receives the percentage adjustment from the strategy optimization decision unit and achieves precise control through two-stage signal processing. The first stage uses a digital PID algorithm to calculate the pulse width modulation parameters:

[0086]

[0087] Where ΔPWM represents the pulse width adjustment amount, K p K i K d These are the proportional, integral, and derivative coefficients, respectively, and e(t) represents the current opening deviation. The second stage outputs a pulse signal with an adjustable duty cycle via a hardware PWM generator, achieving a resolution of 0.1% and a configurable frequency range of 1-10kHz. The output stage employs optocoupler isolation and power amplifier circuitry, providing a drive capability of 10A / 250VAC, enabling direct control of large-diameter steam regulating valves.

[0088] The frequency converter drive module is designed with a three-level topology, and the core processor uses a TIC2000 series DSP. After receiving the frequency adjustment command, the module first performs safety checks and amplitude limiting to prevent overshoot damage to the equipment. The speed control signal is generated using space vector modulation technology, outputting a three-phase PWM waveform with an adjustable carrier frequency of 4kHz. The power unit uses an IPM intelligent module, integrating IGBTs and drive protection circuits, supporting a frequency output range of 0-400Hz. The module has a built-in automatic motor parameter recognition function, making it compatible with solvent circulation pumps of different power ratings.

[0089] The displacement sensing and acquisition module achieves non-contact measurement based on the magnetostrictive principle. The sensor probe is mounted beside the valve stem, with a measurement range of 0-100 mm and a linearity of ±0.05% FS. The signal conditioning circuit includes an instrumentation amplifier and a 24-bit Σ-Δ ADC with a sampling rate of 1 kHz. Data processing employs a sliding median filter combined with a Kalman prediction algorithm to effectively suppress noise caused by mechanical vibration. Displacement data is transmitted via an RS-485 interface using the Modbus RTU protocol with a refresh period of 10 ms.

[0090] The speed feedback acquisition module employs a dual-channel redundant design. The main channel uses a Hall effect sensor with a 60-tooth encoder disk mounted on the motor shaft, generating 60 pulse signals per revolution. The signal is shaped by a Schmitt trigger and then fed into a 32-bit high-speed counter. The backup channel uses vibration spectrum analysis, acquiring the motor housing vibration signal through an accelerometer and extracting the rotational frequency component through FFT transformation. The two data streams are weighted and averaged in a data fusion processor, resulting in a final speed measurement accuracy of ±0.1 rpm.

[0091] The closed-loop verification module runs on an independent real-time operating system partition, with its verification cycle synchronized with the strategy decision window. The module establishes a steam consumption calculation model, with input parameters including valve opening, steam pressure, temperature, and flow feedback data. The model employs a piecewise linear interpolation method, discretizing the equipment characteristic curve into 1000 operating points for table lookup calculation. The deviation detection algorithm incorporates a sliding window statistical mechanism; when the average deviation over three consecutive cycles exceeds a threshold, a feature re-extraction process is triggered. Verification results are transmitted to other modules via shared memory, with a data update latency of less than 5ms.

[0092] For signal transmission, control commands and feedback data use different physical channels. Control signals are transmitted via dedicated shielded twisted-pair cables with an impedance matching of 120Ω and a transmission rate of 1Mbps. Feedback data is transmitted via industrial Ethernet using the IEEE 802.1AS time-sensitive networking protocol to ensure timing determinism. All interface circuits are equipped with TVS diodes and ferrite core filters to meet the IEC 61000-4 electromagnetic compatibility standard.

[0093] The unit incorporates internal fault diagnosis and tolerance mechanisms. At the hardware level, a watchdog circuit and heartbeat detection are employed, automatically switching to a safe mode upon detecting anomalies. At the software level, data validity checks are implemented, including range verification, rate of change limits, and correlation analysis. When sensor failure is detected, the system can switch to a model-based virtual sensing mode to maintain basic operation. Fault events are recorded in non-volatile memory for post-event analysis.

[0094] The human-machine interface is designed as a web service with a built-in lightweight HTTP server. Operators can access real-time data dashboards via a browser to view signal waveforms, device status, and historical trends for each channel. The interface supports online parameter modification; all changes require double confirmation and are logged. Access control uses the RBAC model and supports three-level permission management.

[0095] In terms of environmental adaptability, the unit casing is made of aluminum alloy with an IP65 protection rating. The internal key circuit boards are coated with conformal coating, allowing for condensation resistance up to 95% RH. The operating temperature range is -40℃ to +70℃, achieved through natural cooling via thermal pads and heat sinks. Vibration testing complies with IEC60068-2-6 standards, and it can withstand random vibrations from 5-500Hz.

[0096] This unit ensures the accurate execution of the control strategy through precise signal conversion and a reliable feedback mechanism. Multi-level verification and fault-tolerant design guarantee stable system operation under various conditions, providing a solid foundation for the entire energy optimization system. Real-time data acquisition and processing capabilities make the system state observable, creating the necessary conditions for closed-loop optimization.

[0097] Example 4: See Figure 4 The solvent state analysis unit, as a crucial component of the system, focuses on monitoring and evaluating solvent chemical stability, providing key parameter support for energy consumption optimization. This unit analyzes solvent degradation in real time, outputting chemical stability indicators that directly influence the generation process of control strategies.

[0098] In industrial field implementation, this unit is deployed using a distributed architecture, comprising a solvent composition analysis module and a stability assessment module. The hardware platform uses an Advantech UNO-2484G industrial PC, equipped with an Intel Core i7-8665UE processor and 16GB of memory, running Windows 10 IoT Enterprise. The data acquisition interface supports 4-20mA analog input and Modbus TCP protocol, enabling direct communication with the on-site gas chromatograph.

[0099] The solvent composition analysis module establishes a database of typical degradation products, containing characteristic parameters of 20 common amine solvent degradation products. After receiving the concentration data of the gas phase components from the top of the column, the module executes a multi-step analysis workflow. First, data preprocessing is performed, including baseline correction, peak identification, and noise filtering. Then, retention time locking technology is used to match the detected chromatographic peaks with standards in the database. Finally, the relative content of each component is calculated using peak area normalization, and a concentration distribution report is output. The analysis cycle is controlled within 5 minutes to meet real-time requirements.

[0100] The stability assessment module uses a weighted integral algorithm to process solvent state data. The module has a built-in solvent molecular structure database that records the bond energy parameters of key chemical bonds. The assessment process comprehensively considers the dynamic effects of solvent circulation flow rate and degradation product concentration to calculate a chemical stability index. This index is expressed as a percentage, with 100% representing virgin solvent and 0% representing complete failure. The assessment results are updated every 10 minutes and transmitted to the dynamic energy consumption feature extraction unit via the OPCUA interface.

[0101] The main degradation products and their characteristics monitored by the solvent state analysis unit in a typical MEA solvent system are shown in the table below:

[0102]

[0103] The data acquisition system features a dual-channel redundancy design. The main channel connects to an online gas chromatograph, while the backup channel receives offline analysis data from the laboratory. In the event of a failure in the main channel, the system automatically switches to the backup channel and simultaneously activates data prediction mode to maintain operation. All raw data is stored in an SQLite database with a 90-day retention period, supporting historical data backtesting analysis.

[0104] The chemical stability index is calculated using a multi-factor weighted algorithm. The algorithm inputs include: current solvent circulation flow rate, concentration of each degradation product, cumulative running time, and historical temperature data. A time decay factor is incorporated into the calculation process to reflect the gradual degradation trend of solvent performance. The output results are filtered using a moving average to eliminate short-term fluctuations. The system automatically generates a status report every 5% change in the index value, prompting operators to monitor the solvent condition.

[0105] The human-machine interface adopts a B / S architecture and supports multi-terminal access. The main display areas include real-time chromatograms, degradation product concentration trend graphs, and a chemical stability index dashboard. Operators can view solvent state change curves for any time period and set warning thresholds. When the index falls below the preset limit, the interface automatically displays a warning message and triggers an audible and visual alarm.

[0106] The unit is equipped with self-diagnostic functions to periodically check sensor status and analyze algorithm performance. Diagnostic items include: carrier gas pressure monitoring, detector sensitivity testing, and database integrity verification. When problems are detected, the system handles them according to their severity, ranging from simple warnings to emergency shutdown recommendations. All diagnostic records are written to the system log for maintenance personnel to reference.

[0107] Regarding communication security, the unit implements multiple protection measures. Data transmission uses AES-256 encryption, and access control is based on certificate authentication. Firewall rules are configured on network interfaces to allow only authorized devices to communicate. Modification of critical parameters requires dual authentication, and operation logs are permanently stored and cannot be tampered with.

[0108] On-site installation takes into account actual working conditions. The cabinet is made of 304 stainless steel with an IP54 protection rating, making it suitable for installation in environments with corrosive gases. The internal equipment layout provides sufficient space for heat dissipation, and the ambient temperature is controlled within the range of 0-50℃. The power input is equipped with surge protectors and an isolation transformer to prevent power grid fluctuations from affecting equipment operation.

[0109] The maintenance strategy combines preventative maintenance with condition monitoring. Columns and injection pads are replaced regularly, and detector sensitivity is calibrated monthly. The software is automatically updated, downloading the latest analytical method library via a secure channel. Maintenance records are managed electronically, automatically generating reminders for the next maintenance session.

[0110] Example 5: The multi-channel heat exchange controller, as the core component of the system's energy distribution, is responsible for optimizing the steam flow distribution in each heat exchange branch of the regeneration tower, achieving efficient utilization of thermal energy. Based on the principle of thermodynamic balance, this controller dynamically calculates the flow demand of each branch, generates precise distribution coefficients, and guides the actuators to complete energy allocation.

[0111] In terms of hardware architecture, the controller adopts a distributed I / O module design, with the main control unit being a Beckhoff CX2040 embedded PC equipped with an Intel Atom quad-core processor and 2GB of memory. The expansion modules include 8 channels of analog input, 16 channels of digital output, and 4 channels of RS-485 communication interfaces, meeting the needs of multi-channel signal acquisition and control. The system power supply employs a redundant design, with dual 24VDC inputs and a supercapacitor backup power supply to ensure continuous power supply. The system is mounted using DIN rails and has an IP67 protection rating, making it suitable for harsh industrial environments.

[0112] After receiving the target steam flow rate value from the regeneration strategy optimization decision unit, the controller initiates a multi-stage calculation process. The first stage involves data preprocessing, including validity verification and moving average filtering of the input parameters to eliminate measurement noise interference. The second stage invokes a thermodynamic model, dividing the regeneration tower into three heat exchange zones (upper, middle, and lower), establishing an independent energy balance equation for each zone. Model parameters include real-time data such as lean and rich liquid inlet temperatures, solvent flow rates, and steam saturation temperatures, with the heat transfer coefficient dynamically updated via an online identification algorithm. The third stage performs flow distribution calculations, employing an iterative method to determine the optimal flow ratio for each branch. The calculation results are then output to the actuator after amplitude limiting.

[0113] During the establishment of the thermodynamic model, the controller considers various practical operating conditions. To address solvent viscosity variations, a Reynolds number correction factor is introduced to dynamically adjust the theoretical flow rate ratio. For heat exchanger fouling, the model incorporates a performance degradation curve, automatically compensating for heat transfer efficiency losses based on operating time. Under abnormal operating conditions, such as a faulty temperature sensor in a branch circuit, the system automatically switches to an estimation mode based on the pressure-flow relationship, maintaining basic control functions.

[0114] The control signal output employs a tiered regulation strategy. The main steam pipeline maintains continuous regulation, achieving precise flow control through high-precision electric regulating valves. Each branch uses a discrete regulation method, equipped with multi-position electric ball valves that automatically select the optimal opening combination based on the distribution coefficient. The signal output module integrates opto-isolation and surge protection functions, with an output current range of 4-20mA and a load capacity of up to 500Ω. All control commands are accompanied by timestamps and checksums to ensure reliable transmission.

[0115] The condition monitoring system collects real-time operating parameters for each branch, including steam flow, inlet and outlet temperatures, and pressure. Monitoring data is transmitted via PROFIBUS-DP bus with a sampling period of 100ms. An anomaly detection algorithm monitors the changing trends of each parameter; when leaks, blockages, or other anomalies are detected, a diagnostic report is automatically generated and the allocation strategy is adjusted. Monitoring data is stored on a built-in SSD, using a circular storage method to retain records for the most recent 30 days.

[0116] The communication interface design supports multiple industrial protocols. Interaction with upper-layer systems utilizes the OPCUA protocol with a data transmission rate of 10Mbps. Field device communication supports both Modbus RTU and PROFIBUS modes, flexibly configurable via DIP switches. The network layer implements traffic shaping and priority management to ensure real-time control commands. All communication ports are equipped with electromagnetic isolation protection, conforming to the IEC61000-4-4 standard.

[0117] The human-machine interface is implemented through a 7-inch touchscreen with a resolution of 800×480, supporting gesture operation. The interface is divided into a status monitoring area, a parameter setting area, and an alarm information area. The monitoring area dynamically displays the steam distribution ratio and temperature change curves of each branch; the parameter setting area allows authorized personnel to adjust model coefficients and control parameters; the alarm information area displays current alarms and historical events in categories. The interface design conforms to ergonomic principles, key parameters are distinguished by color, and the operation response time is less than 200ms.

[0118] The controller's safety protection features include multi-level protection measures. At the hardware level, an emergency stop button is provided to directly cut off the power to the actuators. At the software level, multiple interlocking logics are implemented to prevent equipment damage due to misoperation. The safety circuit meets SIL2 requirements and achieves hard-wired protection through independent safety relays. All safety events are recorded in non-volatile memory, supporting fault tracing and analysis.

[0119] In terms of environmental adaptability, the controller operates within a temperature range of -25℃ to +60℃ and has a storage temperature range of -40℃ to +85℃. The housing is made of aluminum alloy with an anodized finish, providing excellent corrosion resistance. The heat dissipation design utilizes natural convection, and an internal temperature sensor monitors the temperature rise of key components in real time. Vibration tolerance meets the IEC60068-2-6 standard, capable of withstanding mechanical vibrations within a frequency range of 5Hz to 150Hz.

[0120] The maintenance functions are designed with on-site requirements in mind. The controller has a built-in self-test program that can periodically perform sensor calibration and actuator testing. The maintenance reminder function automatically generates maintenance plans based on equipment uptime. Firmware supports remote upgrades, and digital signature verification ensures program integrity. The maintenance log records detailed operation records and equipment status changes, which can be exported for analysis via USB interface.

[0121] The implementation of this controller enables more precise and rational heat distribution in the regeneration tower. By dynamically adjusting the steam flow rate of each heat exchange branch, overall thermal efficiency is optimized while avoiding localized overheating or energy waste. Working in conjunction with other modules of the system, it forms a complete control chain from energy consumption strategy generation to heat distribution execution, providing crucial support for the stable and efficient operation of the carbon capture system. Its modular design and strong environmental adaptability ensure reliable performance in various industrial environments.

[0122] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0123] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A real-time energy consumption control system for carbon capture solvent regeneration based on AI optimization, characterized in that, include: The solvent regeneration environment monitoring unit is used to collect multi-dimensional operating parameters in the carbon capture solvent regeneration tower in real time. The multi-dimensional operating parameters include at least the reboiler steam pressure data, solvent circulation flow rate data, lean and rich liquid heat exchange temperature difference data, and tower top gas phase component concentration data. The dynamic energy consumption feature extraction unit is used to receive the multi-dimensional operation parameters and extract the implicit energy consumption feature vector of the solvent regeneration process through a deep spatiotemporal convolutional neural network, while generating a solvent regeneration energy efficiency evaluation index. The regeneration strategy optimization decision unit is used to receive the implicit energy consumption feature vector and the solvent regeneration energy efficiency evaluation index, and to use a reinforcement learning framework to iteratively generate a steam valve opening adjustment strategy and a solvent circulation pump frequency control strategy online. The execution feedback control unit is used to convert the steam valve opening adjustment strategy into a pulse control signal for the steam regulating valve, and at the same time convert the solvent circulation pump frequency control strategy into a speed control signal for the variable frequency drive, and to acquire valve displacement feedback data and pump speed feedback data in real time. The closed-loop verification module is used to compare the deviation between the actual steam consumption data and the theoretical steam consumption data during the verification period. When the deviation exceeds the preset tolerance threshold, the dynamic energy consumption feature extraction unit is triggered to re-extract the implicit energy consumption feature vector of the solvent regeneration process.

2. The AI-optimized real-time energy consumption control system for carbon capture solvent regeneration according to claim 1, characterized in that, The dynamic energy consumption feature extraction unit specifically includes: The parallel convolution processing module receives the reboiler steam pressure time series data and solvent circulation flow time series data transmitted by the solvent regeneration environment monitoring unit, and extracts the steam pressure fluctuation feature map and flow fluctuation feature map respectively through a dual-channel one-dimensional convolution layer. The spatiotemporal correlation analysis module receives the steam pressure fluctuation feature map and flow fluctuation feature map output by the parallel convolution processing module, and uses a three-dimensional convolution kernel to fuse the spatiotemporal dimension correlation features to output a steam-flow coupling feature tensor. The energy efficiency assessment generation module receives the steam-flow coupling characteristic tensor transmitted by the spatiotemporal correlation analysis module, combines the lean and rich liquid heat exchange temperature difference data with the column top gas phase component concentration data, and generates solvent regeneration energy efficiency assessment indicators through full-connected layer regression.

3. The AI-optimized real-time energy consumption control system for carbon capture solvent regeneration according to claim 2, characterized in that, The regeneration strategy optimization decision-making unit performs the following operations: Establish a strategy decision-making time window and obtain the solvent regeneration energy efficiency evaluation index output by the dynamic energy consumption feature extraction unit at a fixed sampling period; The solvent regeneration energy efficiency evaluation index in the current time window and the solvent regeneration energy efficiency evaluation index in the historical time window are input into the long short-term memory network to predict the optimal energy consumption reduction gradient for the next three operating cycles. Based on the predicted optimal energy consumption descent gradient, a dual-action output strategy network is used to generate the percentage change of the steam valve opening adjustment strategy and the frequency adjustment of the solvent circulation pump frequency control strategy.

4. The AI-optimized real-time energy consumption control system for carbon capture solvent regeneration according to claim 3, characterized in that, The execution feedback control unit includes: The valve signal conversion module receives the percentage change of the steam valve opening adjustment strategy transmitted by the regeneration strategy optimization decision unit, and generates a steam regulating valve pulse control signal with a corresponding duty cycle through the pulse width modulation circuit. The variable frequency drive module receives the frequency adjustment amount of the solvent circulation pump frequency control strategy transmitted by the regeneration strategy optimization decision unit, and generates a stepped speed control signal for the variable frequency drive through a digital signal processor. The displacement sensing and acquisition module monitors the valve stem displacement of the steam regulating valve in real time and converts it into valve execution displacement feedback data. The speed feedback acquisition module captures the real-time speed value of the solvent circulation pump motor through a Hall sensor and converts it into pump speed feedback data.

5. The AI-optimized real-time energy consumption control system for carbon capture solvent regeneration according to claim 4, characterized in that, The operation of the closed-loop verification module includes: Obtain historical data on the opening degree of the steam regulating valve and the operating frequency of the solvent circulation pump recorded by the execution feedback control unit; The system calls upon the implicit energy consumption feature vectors for the corresponding time period stored in the dynamic energy consumption feature extraction unit, and calculates the theoretical steam consumption data through the forward propagation network. The absolute deviation between the actual steam consumption data and the theoretical steam consumption data is compared. When the absolute deviation exceeds the preset tolerance threshold for three sampling cycles, a feature re-extraction instruction is sent to the dynamic energy consumption feature extraction unit.

6. The AI-optimized real-time energy consumption control system for carbon capture solvent regeneration according to claim 1, characterized in that, Also includes: The solvent state analysis unit is connected between the solvent regeneration environment monitoring unit and the dynamic energy consumption feature extraction unit. It is used to receive solvent circulation flow data and column top gas phase component concentration data, and output solvent chemical stability index through solvent degradation degree analysis model. The dynamic energy consumption feature extraction unit uses solvent chemical stability index as an additional feature dimension, and inputs it into the deep spatiotemporal convolutional neural network along with multi-dimensional operating parameters.

7. The AI-optimized real-time energy consumption control system for carbon capture solvent regeneration according to claim 6, characterized in that, The solvent state analysis unit includes: The solvent composition analysis module receives the concentration data of gas phase components at the top of the column and analyzes the concentration distribution of solvent degradation products through a gas chromatography simulation algorithm. The stability assessment module receives data on the concentration distribution of solvent degradation products and solvent circulation flow rate, and generates solvent chemical stability indicators based on preset solvent chemical bond breaking thresholds.

8. The AI-optimized real-time energy consumption control system for carbon capture solvent regeneration according to claim 1, characterized in that, The system also includes: A multi-channel heat exchange controller is connected between the regeneration strategy optimization decision unit and the execution feedback control unit. It is used to receive the steam valve opening adjustment strategy and the solvent circulation pump frequency control strategy, and calculate the flow distribution coefficient of each heat exchange branch through the thermodynamic balance model. The execution feedback control unit corrects the amplitude of the steam regulating valve pulse control signal according to the flow distribution coefficient.

9. The AI-optimized real-time energy consumption control system for carbon capture solvent regeneration according to claim 8, characterized in that, The multi-channel heat exchange controller performs the following: Receive the target steam flow value of the current steam valve opening regulation strategy transmitted by the regeneration strategy optimization decision unit; The heat exchange temperature difference data of lean and rich liquids collected by the solvent regeneration environment monitoring unit is obtained, and the theoretical flow rate ratio of each heat exchange branch is calculated by the heat transfer coefficient compensation algorithm. The flow distribution coefficient for each heat exchange branch is generated by multiplying the target steam flow rate by the theoretical flow rate ratio.

10. A method for real-time control of energy consumption for carbon capture solvent regeneration based on AI optimization, characterized in that, Applied to the system as described in any one of claims 1 to 9, the method comprises: The solvent regeneration environment monitoring unit captures real-time data on reboiler steam pressure, solvent circulation flow rate, lean and rich liquid heat exchange temperature difference, and column top vapor phase component concentration. The captured multi-dimensional operation parameters are input into the dynamic energy consumption feature extraction unit, and the hidden energy consumption feature vector and solvent regeneration energy efficiency evaluation index are generated by using a deep spatiotemporal convolutional neural network. A reinforcement learning framework is used to process the implicit energy consumption feature vector and the solvent regeneration energy efficiency evaluation index, and the steam valve opening adjustment strategy and solvent circulation pump frequency control strategy are output online iteratively. The steam valve opening regulation strategy is converted into a steam regulating valve pulse control signal by executing the feedback control unit, and the solvent circulation pump frequency control strategy is converted into a variable frequency drive speed control signal. Real-time acquisition of valve displacement feedback data and pump speed feedback data, combined with theoretical steam consumption data, to perform a closed-loop verification process; When the deviation between the actual steam consumption data and the theoretical steam consumption data exceeds the preset tolerance threshold, the dynamic energy consumption feature extraction unit is triggered to re-extract the implicit energy consumption feature vector.

Images