Solid particle heat storage online dedusting and anti-caking control system and method
By using a multimodal data processing and cause-effect graph-driven control system, the problems of pulverization and caking in high-temperature solid particle thermal storage devices were solved, enabling online dust removal and anti-caking, improving equipment safety and maintenance efficiency, and reducing economic losses.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUADIAN ZHENGZHOU MECHANICAL DESIGN INST
- Filing Date
- 2026-04-08
- Publication Date
- 2026-07-07
Smart Images
Figure CN122345339A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of physical thermal energy storage technology, specifically to an online dust removal and anti-caking control system and method for solid particle thermal energy storage. Background Technology
[0002] As a core technology for long-term, large-scale energy storage, high-temperature solid particle thermal energy storage utilizes gravity-driven, inexpensive sand or ceramic particles to undergo sensible heat charge-discharge cycles between 20℃ and 1000℃. However, it faces severe physicochemical challenges in actual operation: 1. Severe thermal stress pulverization: Particles are subjected to transient temperature gradients of hundreds of degrees in the heat exchanger, resulting in thermal stress breakage and generating a large amount of micron-sized high-temperature dust. This dust is severely worn by the circulating airflow, damaging the waste heat boiler and turbine blades, and clogging the gas-solid heat exchanger flow channels. 2. High-temperature sintering and caking of thermal storage bed: Under high temperatures above 600℃ and huge static pressure and heavy load of tens of meters high silos, impurities melt and solid-phase atomic diffusion occurs at particle contact points, forming "sintering necks", which leads to large-area caking of the thermal storage bed into blocks, causing fluidization failure and local hot spots; 3. Traditional offline maintenance is costly: Traditional offline screening requires cooling the silos (which takes several weeks), resulting in huge economic losses. Summary of the Invention
[0003] The present invention proposes an online dust removal and anti-caking control system and method for solid particle thermal storage, which can at least solve one of the technical problems in the background art.
[0004] To achieve the above objectives, the present invention adopts the following technical solution: An online dust removal and anti-caking control system for solid particle thermal storage is applied to a high-temperature solid particle thermal storage device with an operating temperature of 600℃-1000℃. The system includes a multimodal sensing layer, a multimodal encoder, a physical information causal graph module, a neural symbol inference engine, and a diffusion model generative control unit. The multimodal sensing layer collects real-time pressure time-series data, acoustic spectrum data, thermodynamic temperature field data, and visual monitoring data of the thermal storage bed. The multimodal encoder adopts a linear complexity coding architecture based on a selective state space model to extract features from multi-source heterogeneous data collected by the multimodal perception layer and output a unified multimodal representation vector. The physical information causal graph module has a built-in hard constraint mechanism for the high-temperature gas-solid two-phase flow convection-diffusion partial differential equation. Based on the unified multimodal characterization vector, it constructs a causal directed acyclic graph of the thermal storage system, eliminates false correlations under extreme conditions, and outputs the causal intervention effects of different intervention actions on dust concentration and caking state. The neural symbolic reasoning engine has a built-in physical rule knowledge base in the field of high-temperature thermal storage, dust removal and anti-caking, and performs adaptive weighted fusion of neural network output and symbolic logic reasoning results to output rigid safety constraint boundaries. The diffusion model generative control unit, based on the causal intervention effect and the unified multimodal representation vector, generates a high-dimensional continuous optimal intervention action sequence within a rigid safety constraint boundary, driving the acoustic cleaning and pulse fluidization actuators to complete online dust removal and anti-caking operations.
[0005] As a preferred embodiment of the online dust removal and anti-caking control system for solid particle thermal storage described in this invention, the multimodal encoder includes a timing and acoustic modal encoder, an acoustic spectrum encoder, a visual modal encoder, and a cross-modal attention fusion module. The temporal and acoustic modal encoder adopts a selective state-space model architecture to process high-frequency pressure temporal sequences and particle collision acoustic temporal sequences, realizes input-dependent adaptive memory and forgetting, and outputs temporal acoustic features; The acoustic spectrum encoder is constructed based on a residual network and is used to extract features from the acoustic time-spectrum map obtained by short-time Fourier transform and output spectral features. The visual modal encoder is built based on a self-attention mechanism and is used to extract features from visual monitoring data acquired by industrial cameras and output visual features. The cross-modal attention fusion module is used to calculate the attention weights among temporal acoustic features, spectral features, and visual features, to complete adaptive weighted fusion, and output a unified multimodal representation vector.
[0006] As a preferred embodiment of the solid particle thermal storage online dust removal and anti-caking control system of the present invention, the physical information causal graph module, when constructing a causal directed acyclic graph, incorporates the residual loss of the high-temperature gas-solid two-phase flow convection-diffusion partial differential equation as a hard constraint term into the objective function of causal structure learning, thereby limiting the physical feasible domain of the causal graph search; and after learning is completed, the causal chain is forcibly corrected by preset physical prior rules, the preset physical prior rules including the direct causal influence of temperature on dust generation and caking state, the causal influence of dust concentration on bed pressure, and the causal influence of acoustic excitation parameters on caking state.
[0007] As a preferred embodiment of the solid particle thermal storage online dust removal and anti-caking control system of the present invention, the neural symbol reasoning engine includes an industrial symbol knowledge base and a meta-learning dynamic fusion module. The industrial symbol knowledge base stores the entity set, the semantic relationship set between entities, and the physical rule set of the dust removal and anti-caking system. The physical rule set includes high-temperature caking risk triggering rules, acoustic resonance effect triggering rules, and dust removal efficiency constraint rules. The meta-learning dynamic fusion module adopts a gradient-based two-layer optimized meta-learning algorithm, which adaptively calculates the dynamic fusion weights of neural network output actions and symbolic logic reasoning actions according to the current operating conditions of the system. When abnormal operating conditions such as bed temperature > 800℃ or bed pressure drop > 5.0kPa are detected, the fusion weights are forcibly switched to the symbolic logic-dominated mode to achieve rigid safety cutoff.
[0008] As a preferred embodiment of the solid particle thermal storage online dust removal and anti-caking control system of the present invention, the diffusion model generative control unit adopts a diffusion reverse generation architecture guided by a causal Q function. Taking the current system state and causal intervention effect as conditions, it completes the noise reduction generation through a noise prediction network and outputs a high-dimensional continuous intervention action sequence. Its strategy optimization objective function integrates the score matching term and action regularization term guided by the causal Q function, so that the generated intervention action simultaneously meets the requirements of optimal dust removal and anti-caking effect and safe and stable execution process.
[0009] As a preferred embodiment of the solid particle thermal energy storage online dust removal and anti-caking control system of the present invention, the system further includes a federated learning module, which comprises a global server and multiple local clients corresponding to different thermal energy storage power plants. The local clients are used to update model parameters based on local datasets and add truncated Gaussian noise before uploading parameters to achieve differential privacy protection. The global server uses a weighted aggregation mechanism with a near-end regularization term to aggregate the parameters uploaded by each local client, generate global model parameters, and distribute them to each local client, thereby realizing cross-power plant knowledge collaborative evolution and privacy protection.
[0010] A method for online dust removal and anti-caking control of solid particle thermal storage, applied to high-temperature solid particle thermal storage devices with operating temperatures of 600℃-1000℃, includes the following steps: S1. Real-time acquisition of pressure time-series data, acoustic spectrum data, thermodynamic temperature field data, and visual monitoring data of the thermal storage bed to achieve time synchronization of multi-source data; S2. A linear complexity coding architecture based on a selective state-space model is adopted to extract features and perform cross-modal fusion on the collected multi-source heterogeneous data, and output a unified multimodal representation vector. S3. Based on the unified multimodal characterization vector, a causal directed acyclic graph with hard constraints on the partial differential equation of convection and diffusion of high-temperature gas-solid two-phase flow is constructed to calculate the causal intervention effect of different intervention actions on dust concentration and caking state. S4. Based on the causal intervention effect and the unified multimodal representation vector, a high-dimensional continuous optimal intervention action sequence is generated within the physical rule safety constraints through a causal-guided diffusion model, driving the acoustic cleaning and pulse fluidization actuators to complete online dust removal and anti-caking operations. S5. Based on the causal directed acyclic graph and physical rule knowledge base, generate an interpretable reasoning chain for the corresponding intervention action and simultaneously complete the safety audit of the control process.
[0011] As a preferred embodiment of the online dust removal and anti-caking control method for solid particle thermal storage described in this invention, in step S3, when calculating the causal intervention effect, the do operator is used to cut off all incoming edges of the intervention variable and complete the forced assignment. The set of confounding factors of the parent node of the intervention variable is identified through a causal directed acyclic graph. Based on the backdoor criterion, the confounding factors are marginalized by integration to eliminate the bias of environmental disturbance on the evaluation of the intervention effect, and the conditional average intervention effect of the intervention action on the caking state is output.
[0012] As a preferred embodiment of the online dust removal and anti-caking control method for solid particle thermal storage described in this invention, in step S4, when generating the optimal intervention action sequence, the dynamic fusion weights of the neural network output actions and symbolic logic reasoning actions are adaptively adjusted through a meta-learning mechanism; under normal operating conditions where the bed temperature is ≤700℃ and the bed pressure drop is ≤4.0kPa, the fusion weights are biased towards neural network dominance; under abnormal operating conditions where the bed temperature is >800℃ or the bed pressure drop is >5.0kPa, a forced switch to a symbolic logic-dominated safety mode is performed, and the intervention actions are rigidly truncated.
[0013] As a preferred embodiment of the online dust removal and anti-caking control method for solid particle thermal storage described in this invention, the method further includes a cross-power station collaborative update step: each local node of the thermal storage power station completes local updates of model parameters based on local operating data, adds differential privacy noise to the updated parameters, and uploads them to the global server; the global server uses an aggregation mechanism with a near-end regularization term to complete parameter aggregation, generates a global optimization model, and distributes it to each local node, thereby achieving cross-power station knowledge transfer and model collaborative evolution without leaking the original operating data.
[0014] The beneficial effects of this invention are: This invention's multimodal large model possesses a deep understanding of complex industrial operating conditions, can automatically identify abnormal patterns and generate diagnostic reports, reducing reliance on experts; it surpasses traditional correlation learning, realizing counterfactual reasoning and optimal intervention strategy search; the neural symbolic reasoning engine generates readable decision chains, meeting industrial safety audit requirements, enhancing user trust, and achieving a transformation from "fault response" to "predictive maintenance". Attached Figure Description
[0015] Figure 1 This is the overall logic block diagram of the solid particle thermal storage online dust removal and anti-caking control system and method of the present invention.
[0016] Figure 2This is a multi-modal encoder core logic block diagram of the solid particle thermal storage online dust removal and anti-caking control system and method of the present invention.
[0017] Figure 3 This is a PINN-Causal intervention core logic block diagram of the solid particle thermal storage online dust removal and anti-caking control system and method of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some embodiments of the present invention, but not all embodiments.
[0019] like Figures 1-3 As shown, an online dust removal and anti-caking control system for solid particle thermal storage is applied to a high-temperature solid particle thermal storage device with an operating temperature of 600℃-1000℃. The system includes a multimodal sensing layer, a multimodal encoder, a physical information causal graph module, a neural symbol inference engine, and a diffusion model generative control unit. The multimodal sensing layer collects real-time pressure time-series data, acoustic spectrum data, thermodynamic temperature field data, and visual monitoring data of the thermal storage bed. The multimodal encoder adopts a linear complexity coding architecture based on a selective state space model to extract features from multi-source heterogeneous data collected by the multimodal perception layer and output a unified multimodal representation vector. The physical information causal graph module has a built-in hard constraint mechanism for the convection and diffusion partial differential equation of high-temperature gas-solid two-phase flow. Based on the unified multimodal characterization vector, it constructs a causal directed acyclic graph of the thermal storage system, eliminates false correlations under extreme conditions, and outputs the causal intervention effects of different intervention actions on dust concentration and caking state. The neural symbolic reasoning engine has a built-in physical rule knowledge base for the field of high-temperature thermal storage, dust removal and anti-caking. It adaptively weights and fuses the neural network output and symbolic logic reasoning results to output rigid safety constraint boundaries. The diffusion model generative control unit, based on causal intervention effects and a unified multimodal representation vector, generates a high-dimensional continuous optimal intervention action sequence within a rigid safety constraint boundary, driving the acoustic cleaning and pulse fluidization actuators to complete online dust removal and anti-caking operations.
[0020] Furthermore, the multimodal encoder includes a temporal and acoustic modal encoder, an acoustic spectrum encoder, a visual modal encoder, and a cross-modal attention fusion module; The temporal and acoustic modal encoder adopts a selective state-space model architecture to process high-frequency pressure temporal sequences and particle collision acoustic temporal sequences, achieving input-dependent adaptive memory and forgetting, and outputting temporal acoustic features; Specifically, a selective state-space model enables ultra-fast multimodal coding with linear complexity:
[0021]
[0022] Discretized via a zero-order hold (ZOH):
[0023] In the formula: Hidden state representation vector; : The dynamic state-space parameters that the input depends on; : Discretized state transition matrix; : The input time series sequence; : Output temporal acoustic features; The acoustic spectrum encoder is built based on a residual network and is used to extract features from the acoustic time-spectrum map obtained by short-time Fourier transform and output spectral features. Specifically, acoustic modal encoders:
[0024] In the formula: The encoded output vector of the acoustic mode, with dimension . ; : The original acoustic signal time-domain waveform; Short-time Fourier transform converts a time-domain signal into a time-spectrum graph. Acoustic feature extraction network based on residual network, used to extract deep features from spectrograms; The visual modal encoder is built on a self-attention mechanism and is used to extract features from visual monitoring data acquired by industrial cameras and output visual features. Specifically, visual modal encoder:
[0025] In the formula: The encoded output vector of the visual modality, with dimensions of ; The current industrial camera image input; A visual feature extraction network based on self-attention mechanism; The cross-modal attention fusion module is used to calculate the attention weights between temporal acoustic features, spectral features, and visual features, to complete adaptive weighted fusion, and output a unified multimodal representation vector; Specifically, the multimodal cross-attention fusion mechanism:
[0026]
[0027] In the formula: Modality For modes Normalized attention weight coefficients; : Encoding representation vectors for each modality; : The projection matrix of query, key, and value; : Scaling factor, i.e. the dimension of the key vector, used to prevent gradient vanishing; : in modality The unified representation vector after incorporating information from other modalities; Exponential function; For all modes Summation is used to achieve normalization; in An adaptive weighted fusion of information from different modalities is achieved through a cross-modal attention mechanism.
[0028] Among them, the physical information causal graph module incorporates the residual loss of the high-temperature gas-solid two-phase flow convection-diffusion partial differential equation as a hard constraint term into the objective function of causal structure learning when constructing the causal directed acyclic graph, thus limiting the physical feasible domain of the causal graph search. After learning, the causal chain is forcibly corrected through preset physical prior rules, which include the direct causal influence of temperature on dust generation and caking state, the causal influence of dust concentration on bed pressure, and the causal influence of acoustic excitation parameters on caking state.
[0029] Specifically, the partial differential equations of high-temperature fluid dynamics are injected as hard constraints into causal graph learning, and structure discovery is performed in conjunction with continuous optimization algorithms:
[0030] Among them, the PDE physical loss term of thermodynamics and dust diffusion Defined as:
[0031] In the formula: : A matrix containing observation data; Weighted adjacency matrix; : Residual loss based on the physical mechanism of convection-diffusion equation; Hidden variables of dust concentration, temperature field, and pressure field; The airflow velocity vector is calculated from the bed pressure gradient and the permeability of the porous medium using Darcy's law. ,in For penetration rate, The gas dynamic viscosity; : Turbulent diffusion coefficient, from - Calculation of the closure equation for the turbulence model: ,in , For turbulent kinetic energy, The dissipation rate; : Pulverization source term, given by the thermal stress fracture model: ,in This is an empirical coefficient; This architecture ensures the generation of causal graphs. Achieving dual convergence in both mathematical and physical mechanisms completely eliminates black-box bias caused by purely data-driven approaches.
[0032] Furthermore, the physical information causal graph module uses the do operator and backdoor criterion to calculate the causal intervention effect. By marginalizing confounding factors, it eliminates the bias of environmental disturbances in the assessment of the intervention effect and outputs the conditional average intervention effect of the intervention action on the compaction state. The calculation formula is as follows:
[0033] In the formula: In a given context In this case, intervention is applied to the acoustic parameters. Afterwards, the solidified state causal probability distribution; The :do operator indicates a cutoff. All incoming edges and force assignment to a value ; Nodes in a cause-and-effect graph The set of parent nodes (i.e., confounding factors); Given the context, the prior distribution of the confounding factors can be used to estimate unbiased causal effects by marginalizing the confounding factors.
[0034] The neural symbolic reasoning engine includes an industrial symbolic knowledge base and a dynamic fusion module for meta-learning; The industrial symbol knowledge base stores the entity set, the semantic relationship set between entities, and the physical rule set of the dust removal and anti-caking system. The physical rule set includes high-temperature caking risk triggering rules, acoustic resonance effect triggering rules, and dust removal efficiency constraint rules. The meta-learning dynamic fusion module adopts a gradient-based two-layer optimized meta-learning algorithm, which adaptively calculates the dynamic fusion weights of the neural network output actions and symbolic logic reasoning actions according to the current operating conditions of the system. When abnormal operating conditions such as bed temperature > 800℃ or bed pressure drop > 5.0kPa are detected, the fusion weights are forcibly switched to the symbolic logic-dominated mode to achieve rigid safety cutoff.
[0035] Specifically, construct a symbolic knowledge base for the field of industrial dust removal and anti-caking. :
[0036] In the formula: Symbolic knowledge base, containing formal representations of domain knowledge; : A collection of entities, containing physical objects and concepts within the system; : Relationship set, defining the semantic associations between entities; A set of rules that contains expert knowledge of condition-action patterns; Physical rule set
[0037] Neural module: responsible for perception and pattern recognition
[0038] In the formula: : Feature vectors output by the neural network module; : Neural network encoder, based on deep learning architecture; :time Multimodal observation input; Symbolic module: responsible for logical reasoning and decision-making.
[0039] In the formula: Action decisions output by the symbolic reasoning module; Symbolic reasoning engine, performs logical reasoning; : Knowledge base query function, which retrieves relevant knowledge based on feature vectors; Symbol knowledge base; : Feature vectors output by the neural network module; Integration mechanism
[0040] In the formula: :time The final action decision; : The probability distribution of actions output by the neural network module; : The action probability distribution output by the symbolic reasoning module; Dynamic fusion weights, value range ; The weights of symbolic reasoning are complementary to the weights of neural networks; in The weights are dynamically fused and adaptively determined by the meta-learner based on the scenario. The quantitative triggering conditions for dynamic fusion weights are shown in Table 1 below: Table 1
[0041] Furthermore,
[0042] In the formula: Dynamic fusion weights The sigmoid activation function maps the output to... interval Weight matrix of meta-learner Feature vectors output by the neural network module Action decisions output by the symbolic reasoning module : Current context state vector Vector concatenation operation.
[0043] Among them, the diffusion model generative control unit adopts a diffusion reverse generation architecture guided by the causal Q function. It uses the current system state and causal intervention effect as conditions, and completes the denoising generation through the noise prediction network to output a high-dimensional continuous intervention action sequence. Its strategy optimization objective function integrates the score matching term and action regularization term guided by the causal Q function, so that the generated intervention action can simultaneously meet the requirements of optimal dust removal and anti-caking effect and safe and stable execution process.
[0044] Specifically, designing a policy network for causal perception. Encoding the causal graph structure into a neural network:
[0045] In the formula: :time Selected control action; : Parameters are The strategy network; :time The state vector; Cause-effect graph structure; : Weight matrix of the action output layer; : Vector concatenation operation, which embeds and concatenates the state vector with the cause-effect graph; Graph convolutional networks are used to encode structural information of causal graphs. : Normalized exponential function, outputs the probability distribution of actions; The strategy optimization objective incorporates score matching guided by a causal Q-function:
[0046] In the formula: The loss function of the Actor policy network; : Action value evaluated by causal Critic network; : The guiding strength coefficient of the causal Q function; Action regression regularization coefficient; Expert demonstration of the action or the best historical action.
[0047] Furthermore, the system also includes a federated learning module, which consists of a global server and multiple local clients corresponding to different thermal storage power plants. The local clients are used to update model parameters based on local datasets and add truncated Gaussian noise before uploading parameters to achieve differential privacy protection. The global server uses a weighted aggregation mechanism with a near-end regularization term to aggregate the parameters uploaded by each local client, generate global model parameters, and distribute them to each local client to achieve cross-power plant knowledge collaborative evolution and privacy protection.
[0048] Specifically, federated learning architecture Global server:
[0049] In the formula: : Global model parameters; Aggregate functions, typically using weighted averages; :all The local model parameter set for each client; The total number of clients participating in federated learning; Local clients: Various energy storage power stations
[0050] In the formula: : No. Local model parameters for each client; : Local update function, performing gradient descent optimization; : Global model parameters, used as initialization for local training; : No. Local datasets for each client; Privacy protection mechanism Differential privacy is used to protect gradient information:
[0051] In the formula: The gradient after adding noise is used to upload it to the server. : No. The original gradients for each client; The mean is 0 and the variance is 0. Gaussian noise; Noise standard deviation, controlling the intensity of privacy protection; : Identity matrix, with the same dimensions as the gradient; Non-independent and identically distributed data processing Design a federated aggregation mechanism with near-end regularization to handle the differences in data distribution among power plants:
[0052] In the formula: : No. Local model parameters for each client; Learning rate: controls the step size for updating parameters; Local loss function with respect to parameters The gradient; Proximal term coefficients control the degree of deviation between the local model and the global model; : Global model parameters; Proximal terms constrain the local model from deviating too far from the global model; On the other hand, the present invention provides a method for online dust removal and anti-caking control of solid particle thermal storage, applied to the above-mentioned system, comprising the following steps: S1. Real-time acquisition of pressure time-series data, acoustic spectrum data, thermodynamic temperature field data, and visual monitoring data of the thermal storage bed, and microsecond-level time synchronization of multi-source data through the IEEE1588PTP protocol; S2. A linear complexity coding architecture based on a selective state-space model is adopted to extract features from the collected multi-source heterogeneous data, and feature fusion is completed through a cross-modal attention mechanism to output a unified multimodal representation vector. S3. Based on the unified multimodal characterization vector, a causal directed acyclic graph with hard constraints on the partial differential equation of convection and diffusion of high-temperature gas-solid two-phase flow is constructed. The causal intervention effect of different intervention actions on dust concentration and caking state is calculated by using the do operator and backdoor criterion. S4. Based on the causal intervention effect and the unified multimodal representation vector, a high-dimensional continuous optimal intervention action sequence is generated within the physical rule safety constraints of the neural symbolic reasoning engine through a causal-guided diffusion model, driving the acoustic cleaning and pulse fluidization actuators to complete online dust removal and anti-caking operations. S5. Based on the causal directed acyclic graph and physical rule knowledge base, generate a natural language interpretable reasoning chain for the corresponding intervention action, and simultaneously complete the safety audit and log retention of the control process.
[0053] Specific implementation examples are as follows: The interaction control steps between the intelligent system and the real physical device of this invention are as follows: Step 1: High-frequency synchronous acquisition and hard synchronization of multimodal physical signals Physical data acquisition: The temperature field, local pressure drop and acoustic spectrum of particle friction and collision in the fluidized bed layer inside the silo are acquired in real time by using a 128-point FBG fiber array (sampling rate 1Hz), 32 high-temperature differential pressure transmitters (sampling rate 100Hz) and a 16-channel ultrasonic microphone (sampling rate 20kHz) arranged on the wall of the thermal storage silo. Clock synchronization: All physical sensor signals are hard-synchronized with microsecond-level timestamps on the edge-side industrial control computer via the IEEE 1588PTP protocol; Step 2: Extraction of fluidized state physical features at the edge computing end Acoustic feature calculation: Edge computing nodes perform continuous short-time Fourier transform (STFT) on the 20kHz microphone signal to capture the 10kHz-15kHz high-frequency acoustic excitation features caused by particle sintering neck fracture or pulverization. Physical resistance assessment: Based on the differential pressure transmitter data, the local resistance coefficient of the gas-solid two-phase flow through the heat exchanger tube bundle is calculated in real time. When a pressure drop gradient is detected... Furthermore, when the local temperature exceeds 750℃, a physical anti-caking warning is triggered.
[0054] Step 3: Solving the intervention command of the causal physics engine State mapping: The collected physical signals are input into the compressed state-space multimodal model, and the output is a three-dimensional topology map of the current silo's slab caving risk. Intervention prediction: The causal engine calculates the physical penetration and dust removal effect (CATE index) of the current dead zone if different sound wave frequencies (such as 75Hz vs 150Hz) are turned on, based on the Navier-Stokes convection equation. Step 4: Cooperative control and closed-loop operation of multiple physical actuators Command issuance: The system issues an intervention command to the PLC control cabinet: First, the low-frequency high-intensity pneumatic acoustic wave cleaner at 75Hz and 145dB is continuously vibrated for 3 minutes to break the solid phase sintering neck between particles; then, the nitrogen pulse valve at 0.6MPa is triggered to perform a short backflushing for 50 milliseconds. Physical closed-loop verification: After the intervention is executed, if the differential pressure transmitter detects that the local pressure drop falls back to the design reference value of 2.5 kPa within 10 seconds, the physical intervention is deemed successful and a closed-loop control is completed. Experimental Results and Physical Performance Verification After deploying this system, a three-month continuous operation test was conducted on a 10MW / 40MWh thermal energy storage demonstration project, and the following significant improvements in industrial physical indicators were achieved, as shown in Tables 2, 3, and 4: Table 2 Physical heat transfer performance and flow resistance recovery
[0055] Table 3 Physical Indicators of Particle Powdering and Dust Removal
[0056] Table 4 Response Delay of Hardware Control System
[0057] It should be noted that, in this document, 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 a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0058] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0059] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A solid particle heat storage online dedusting and anti-caking control system applied to a high-temperature solid particle heat storage device with a working temperature of 600-1000℃, characterized in that, It includes a multimodal perception layer, a multimodal encoder, a physical information causal graph module, a neural symbolic reasoning engine, and a diffusion model generative control unit; The multimodal sensing layer collects real-time pressure time-series data, acoustic spectrum data, thermodynamic temperature field data, and visual monitoring data of the thermal storage bed. The multimodal encoder adopts a linear complexity coding architecture based on a selective state space model to extract features from multi-source heterogeneous data collected by the multimodal perception layer and output a unified multimodal representation vector. The physical information causal graph module has a built-in hard constraint mechanism for the high-temperature gas-solid two-phase flow convection-diffusion partial differential equation. Based on the unified multimodal characterization vector, it constructs a causal directed acyclic graph of the thermal storage system, eliminates false correlations under extreme conditions, and outputs the causal intervention effects of different intervention actions on dust concentration and caking state. The neural symbolic reasoning engine has a built-in physical rule knowledge base in the field of high-temperature thermal storage, dust removal and anti-caking, and performs adaptive weighted fusion of neural network output and symbolic logic reasoning results to output rigid safety constraint boundaries. The diffusion model generative control unit, based on the causal intervention effect and the unified multimodal representation vector, generates a high-dimensional continuous optimal intervention action sequence within a rigid safety constraint boundary, driving the acoustic cleaning and pulse fluidization actuators to complete online dust removal and anti-caking operations.
2. The online dust removal and anti-caking control system for solid particle thermal storage according to claim 1, characterized in that: The multimodal encoder includes a temporal and acoustic modal encoder, an acoustic spectrum encoder, a visual modal encoder, and a cross-modal attention fusion module; The temporal and acoustic modal encoder adopts a selective state-space model architecture to process high-frequency pressure temporal sequences and particle collision acoustic temporal sequences, realizes input-dependent adaptive memory and forgetting, and outputs temporal acoustic features; The acoustic spectrum encoder is constructed based on a residual network and is used to extract features from the acoustic time-spectrum map obtained by short-time Fourier transform and output spectral features. The visual modal encoder is built based on a self-attention mechanism and is used to extract features from visual monitoring data acquired by industrial cameras and output visual features. The cross-modal attention fusion module is used to calculate the attention weights among temporal acoustic features, spectral features, and visual features, to complete adaptive weighted fusion, and output a unified multimodal representation vector.
3. The online dust removal and anti-caking control system for solid particle thermal storage according to claim 1, characterized in that: When constructing a causal directed acyclic graph, the physical information causal graph module incorporates the residual loss of the high-temperature gas-solid two-phase flow convection-diffusion partial differential equation as a hard constraint term into the objective function of causal structure learning, thus limiting the physical feasible domain of the causal graph search. After learning, the causal chain is forcibly corrected through preset physical prior rules, which include the direct causal influence of temperature on dust generation and caking state, the causal influence of dust concentration on bed pressure, and the causal influence of acoustic excitation parameters on caking state.
4. The online dust removal and anti-caking control system for solid particle thermal storage according to claim 1, characterized in that: The neural symbolic reasoning engine includes an industrial symbolic knowledge base and a dynamic fusion module for meta-learning. The industrial symbol knowledge base stores the entity set, the semantic relationship set between entities, and the physical rule set of the dust removal and anti-caking system. The physical rule set includes high-temperature caking risk triggering rules, acoustic resonance effect triggering rules, and dust removal efficiency constraint rules. The meta-learning dynamic fusion module adopts a gradient-based two-layer optimized meta-learning algorithm, which adaptively calculates the dynamic fusion weights of neural network output actions and symbolic logic reasoning actions according to the current operating conditions of the system. When abnormal operating conditions such as bed temperature > 800℃ or bed pressure drop > 5.0kPa are detected, the fusion weights are forcibly switched to the symbolic logic-dominated mode to achieve rigid safety cutoff.
5. The online dust removal and anti-caking control system for solid particle thermal storage according to claim 1, characterized in that: The diffusion model generative control unit adopts a diffusion reverse generation architecture guided by the causal Q function. It uses the current system state and causal intervention effect as conditions, and completes the denoising generation through a noise prediction network to output a high-dimensional continuous intervention action sequence. Its strategy optimizes the objective function by integrating score matching terms and action regularization terms guided by the causal Q function, so that the generated intervention actions simultaneously meet the requirements of optimal dust removal and anti-caking effect and safe and stable execution process.
6. The online dust removal and anti-caking control system for solid particle thermal storage according to claim 1, characterized in that: The system also includes a federated learning module, which comprises a global server and multiple local clients corresponding to different thermal power plants. The local clients are used to update model parameters based on local datasets and add truncated Gaussian noise before uploading parameters to achieve differential privacy protection. The global server uses a weighted aggregation mechanism with a near-end regularization term to aggregate the parameters uploaded by each local client, generate global model parameters, and distribute them to each local client, thereby achieving cross-power plant knowledge collaborative evolution and privacy protection.
7. A method for online dust removal and anti-caking control of solid particle thermal storage, applied to high-temperature solid particle thermal storage devices with an operating temperature of 600℃-1000℃, characterized in that, Includes the following steps: S1. Real-time acquisition of pressure time-series data, acoustic spectrum data, thermodynamic temperature field data, and visual monitoring data of the thermal storage bed to achieve time synchronization of multi-source data; S2. A linear complexity coding architecture based on a selective state-space model is adopted to extract features and perform cross-modal fusion on the collected multi-source heterogeneous data, and output a unified multimodal representation vector. S3. Based on the unified multimodal characterization vector, a causal directed acyclic graph with hard constraints on the partial differential equation of convection and diffusion of high-temperature gas-solid two-phase flow is constructed to calculate the causal intervention effect of different intervention actions on dust concentration and caking state. S4. Based on the causal intervention effect and the unified multimodal representation vector, a high-dimensional continuous optimal intervention action sequence is generated within the physical rule safety constraints through a causal-guided diffusion model, driving the acoustic cleaning and pulse fluidization actuators to complete online dust removal and anti-caking operations. S5. Based on the causal directed acyclic graph and physical rule knowledge base, generate an interpretable reasoning chain for the corresponding intervention action and simultaneously complete the safety audit of the control process.
8. The method for online dust removal and anti-caking control of solid particle thermal storage according to claim 7, characterized in that: In step S3, when calculating the causal intervention effect, the do operator is used to cut off all incoming edges of the intervention variable and complete the forced assignment. The set of confounding factors of the parent node of the intervention variable is identified through the causal directed acyclic graph. Based on the backdoor criterion, the confounding factors are marginalized by integration to eliminate the bias of environmental disturbance on the evaluation of the intervention effect. The conditional average intervention effect of the intervention action on the compaction state is output.
9. The method for online dust removal and anti-caking control of solid particle thermal storage according to claim 7, characterized in that: In step S4, when generating the optimal intervention action sequence, the dynamic fusion weights of the neural network output action and the symbolic logic reasoning action are adaptively adjusted through a meta-learning mechanism. Under normal operating conditions, when the bed temperature is ≤700℃ and the bed pressure drop is ≤4.0kPa, the fusion weights are biased towards neural network dominance. Under abnormal operating conditions, when the bed temperature is >800℃ or the bed pressure drop is >5.0kPa, the system is forcibly switched to a safe mode dominated by symbolic logic, and the intervention action is rigidly truncated.
10. The method for online dust removal and anti-caking control of solid particle thermal storage according to claim 7, characterized in that: It also includes a cross-power station collaborative update step: the local nodes of each thermal energy storage power station complete the local update of model parameters based on local operating data, add differential privacy noise to the updated parameters and then upload them to the global server; The global server uses an aggregation mechanism with near-end regularization to aggregate parameters, generate a global optimization model, and distribute it to each local node, achieving cross-power station knowledge transfer and model co-evolution without leaking the original operating data.