An intelligent processing system and processing method for energy conversion

Abstract

This invention relates to an intelligent processing system and method for energy conversion, belonging to the field of energy conversion and control technology. It involves real-time acquisition of operational data through a sensor array; edge computing nodes preprocessing data using a sliding window averaging method and a cross-correlation delay estimation method to extract equipment health, energy output efficiency, and environmental coupling characteristics; a digital twin model employing a mechanistic model and an LSTM error correction model in parallel computation to output a sequence of predicted future states; a central decision server using a deep deterministic policy gradient algorithm to solve for the optimal control action with a multi-objective function encompassing efficiency, safety, and cost; a coordinating controller issuing commands to adjust the actuators; and repeating the above steps to form a closed-loop control. This invention reduces system response time and improves energy conversion efficiency through a collaborative architecture of digital twin prediction, reinforcement learning decision-making, and knowledge graph enhancement.

Description

Technical Field

[0001] This invention relates to the field of energy conversion and control technology, specifically to an intelligent processing system and method for energy conversion. Background Technology

[0002] Energy conversion technology is a crucial pathway to achieving cascaded energy utilization and energy conservation and emission reduction. Technologies such as natural gas pressure power generation, LNG cold energy recovery, and waste heat power generation significantly improve energy efficiency by converting primary energy sources (pressure energy, cold energy, and heat energy) into electrical or thermal energy. Taking natural gas pressure power generation as an example, high-pressure natural gas expands through a turbine expander, driving a generator to produce electricity; a single 500kW system can generate over 3 million kWh annually. However, existing energy conversion systems commonly rely on PID control or manual experience-based operation, which struggles to handle complex issues such as fluctuations in multiple energy inputs, nonlinear time-varying characteristics of equipment, and strong coupling of operating conditions. This results in a significant gap between actual operating efficiency and design efficiency (typically 60%-75%). Furthermore, the lack of real-time assessment methods for equipment health leads to low accuracy in fault warnings, resulting in frequent unplanned shutdowns that cause economic losses and safety risks. Moreover, multiple objectives such as efficiency, safety, and cost often conflict, making it difficult for traditional control strategies to achieve dynamic balance and global optimization.

[0003] To address the aforementioned issues, academia and industry have conducted relevant research in recent years. Digital twin technology, by constructing virtual mappings of physical entities, can achieve online simulation and state prediction of energy conversion processes; deep reinforcement learning algorithms can handle optimization decision-making problems in continuous action spaces, showing potential in the field of multi-objective optimization; knowledge graph technology, by structurally storing the correlation between historical operating conditions and disposal strategies, can provide empirical support for real-time decision-making. However, existing technical solutions mostly focus on the application of single technical paths, lacking a collaborative architecture that organically integrates digital twin prediction, reinforcement learning decision-making, and knowledge graph experience. Specifically, existing digital twin models mostly adopt pure mechanism modeling or pure data-driven modeling. The former is difficult to accurately describe complex nonlinear factors (such as equipment aging and friction loss), while the latter lacks physical constraints, resulting in insufficient generalization ability; the reinforcement learning decision-making process lacks prior knowledge guidance, has slow convergence speed, and insufficient reliability under edge conditions; multi-source heterogeneous sensor data lacks effective preprocessing methods, and temporal inconsistencies and noise interference affect the accuracy of subsequent modeling. In addition, the safety redundancy mechanism of existing systems under abnormal operating conditions is not perfect, making it difficult to meet the high reliability requirements of industrial sites. Summary of the Invention

[0004] In order to solve the problems of the prior art, the present invention provides an intelligent processing system and method for energy conversion.

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: Firstly, an intelligent processing method for energy conversion, comprising the following steps: S1: By deploying sensor groups in each energy conversion unit, operating data is collected in real time at a preset fixed sampling frequency. The operating data includes at least voltage, current, temperature, pressure, and flow rate. S2: The operational data is preprocessed and features are extracted using edge computing nodes to generate a set of state features that characterize the current system operating condition; wherein... The preprocessing includes: smoothing and filtering the data using a sliding window averaging method to eliminate random noise, and aligning the timestamps of sensor data with different sampling frequencies using a cross-correlation delay estimation method. The sliding window averaging method, a classic digital signal processing technique, suppresses high-frequency random noise by calculating the average value of data within a window, resulting in smoother data curves and a cleaner signal for subsequent feature extraction. Simultaneously, cross-correlation delay estimation is used for timestamp alignment because different sensors (such as temperature and pressure sensors) may have different sampling frequencies and response times. By calculating the correlation coefficients of two signals under different delays, the delay value that maximizes the correlation is found, thereby aligning the multi-source data in the time dimension and ensuring spatiotemporal consistency for subsequent analysis.

[0006] The feature extraction includes: generating equipment health features based on the comparison of vibration amplitude and preset alarm threshold; generating energy output efficiency features based on the ratio of output electrical power to input primary energy power; and generating environmental coupling features based on the difference between ambient temperature and standard temperature. Feature extraction is not simply listing raw data, but rather extracting three features with clear physical meaning: the equipment health feature, generated by comparing vibration amplitude with the preset alarm threshold, directly reflects the mechanical state of the equipment; the energy output efficiency feature, generated by the ratio of output electrical power to input power, directly measures the economic efficiency of energy conversion; and the environmental coupling feature, generated by the difference between ambient temperature and standard temperature, reflects the impact of the external environment on system operation. These three features together constitute a state feature set, condensing massive amounts of raw data into key information characterizing the system's operating condition, greatly reducing the computational burden of subsequent models.

[0007] S3: Input the state feature set into the digital twin model, which contains two parallel computing sub-modules: The first submodule is a mechanism model built based on thermodynamic and fluid dynamic equations, used to simulate the physical operating state of the energy conversion unit according to the current state feature set; The second submodule is an error correction model built on a long short-term memory neural network, which is used to predict the output error of the mechanism model at the current moment by taking the prediction errors of the most recent multiple time points as input. The simulation results of the mechanism model are added to the prediction error output by the error correction model to obtain the corrected prediction result, and the predicted state sequence within the future preset time period is output based on this. S4: Input the state feature set and predicted state sequence into a decision model based on deep reinforcement learning. The decision model adopts a deep deterministic policy gradient algorithm, using a multi-objective function containing energy conversion efficiency, safety margin, and operating cost as the reward function. Through the interactive iteration of the policy network and the value network, the optimal control action that maximizes the cumulative reward is solved. S5: Convert the optimal control action into control commands and send them to the actuators of each energy conversion unit to adjust the valve opening, generator speed or power factor; S6: Repeat steps S1 to S5 to form a closed-loop control.

[0008] In one aspect of the first method, the sliding window averaging method is used in S2 to perform smoothing filtering with a window size of N=10. The cross-correlation delay estimation method calculates the correlation coefficient of the two signals under different delays and takes the delay value when the correlation coefficient is the largest as the alignment basis.

[0009] In one aspect of the first method, the long short-term memory neural network in S3 has a 5-layer structure with 128 hidden layer nodes. The input is the error vector between the mechanistic model output and the actual value in the last 5 time steps, and the output is the prediction error value at the current time step.

[0010] In one aspect of the first method, the reward function of the deep deterministic policy gradient algorithm in S4 is specifically as follows: in, To improve overall energy conversion efficiency, As a safety margin indicator, As an operating cost indicator, , , The preset weight coefficients are used; the strategy network is a 3-layer fully connected neural network, and the value network is a 3-layer fully connected neural network.

[0011] One approach in the first aspect also includes exception handling steps: After step S1, the operating data is monitored in real time to see if it exceeds the preset safety threshold. If the temperature exceeds 100°C or the pressure exceeds 1.5MPa, an emergency shutdown is triggered immediately. If communication is interrupted for more than 5 seconds, the local backup controller will be activated and operation will continue according to the preset maintenance strategy.

[0012] In one aspect of the first method, the method further includes a knowledge graph-based decision enhancement step: Before step S4, the current state feature set is matched with the pre-built energy conversion knowledge graph for similarity. Historical operating conditions and their optimal control strategies with a cosine similarity greater than 0.85 are retrieved, and the retrieval results are used as the initial policy or action output constraints of the policy network in the decision model.

[0013] In one aspect of the first method, the performance metrics include: The system response time is less than 100 milliseconds, from data acquisition to command issuance; Energy conversion efficiency is improved by more than 5% compared to a baseline system that does not employ intelligent control; The accuracy rate of equipment fault early warning exceeds 90%, which is verified based on historical fault data.

[0014] Secondly, an intelligent processing system for energy conversion, used to execute processing methods, including: The physical conversion layer contains multiple energy conversion units, each equipped with a sensor array and an actuator; Edge computing nodes, which are communicatively connected to the sensor group, are used to perform preprocessing and feature extraction in step S2; A digital twin server, connected to an edge computing node, is used to perform the simulation and prediction in step S3; The central decision server, connected to the digital twin server, is internally deployed with a deep reinforcement learning model for executing the decision in step S4; The collaborative controller is connected to both the central decision-making server and the execution mechanism, and is used to issue instructions for step S5.

[0015] In one aspect of the second method, the central decision server is also connected to a knowledge graph database, which stores historical operating conditions and corresponding optimal control strategies for decision enhancement.

[0016] In a third aspect, a computer-readable storage medium has a computer program stored thereon, which, when executed by a processor, implements a processing method.

[0017] The beneficial effects of this invention are as follows: 1. By constructing a closed-loop control architecture of "sensing-processing-prediction-decision-execution," intelligent collaborative optimization of the energy conversion process is achieved. First, edge computing nodes use a sliding window averaging method and cross-correlation delay estimation method to preprocess multi-source sensor data in real time, eliminating high-frequency random noise and aligning the time dimension of data from different sampling frequencies. This provides a high signal-to-noise ratio and time-consistent state feature set for subsequent modeling. The digital twin model employs a hybrid modeling strategy of parallel computation of a mechanistic model and an LSTM error correction model: the mechanistic model simulates physical operating states based on the first law of thermodynamics and fluid dynamics control equations, ensuring that the prediction results conform to basic physical laws; the LSTM error correction model uses the prediction errors from multiple recent moments as input to learn and compensate for the systematic biases of the mechanistic model. The high-fidelity prediction result obtained by adding the two models retains the reliability of the physical model while improving prediction accuracy through data-driven correction, enabling the system to have a forward-looking perception capability for future operating conditions. 2. The central decision server employs a deep deterministic policy gradient algorithm, using a multi-objective function encompassing energy conversion efficiency, safety margin, and operating cost as the reward function. It iteratively solves for the optimal control action through the interaction of the policy network and value network, achieving a dynamic balance between efficiency, safety, and cost. A knowledge graph database stores the correlation between historical operating conditions and optimal policies. During online operation, cosine similarity matching is used to retrieve similar historical experiences and apply them as initial policy or action constraints for the decision model, significantly accelerating the convergence process of reinforcement learning and improving decision reliability. The collaborative controller converts decision commands into control signals recognizable by the actuators, adjusting valve opening, generator speed, or power factor in real time. An anomaly handling mechanism monitors key parameters such as temperature and pressure in real time, triggering an emergency shutdown when safety thresholds are exceeded. In the event of communication interruption, a local backup controller is activated to maintain operation according to a preset strategy, ensuring system safety under extreme conditions. The entire system achieves dynamic adaptive adjustment to changes in operating conditions through a closed-loop cycle from S1 to S6, ultimately resulting in a system response time of less than 100 milliseconds, an energy conversion efficiency improvement of more than 5%, and an equipment fault early warning accuracy rate of over 90%, significantly improving the operational economy and safety of the energy conversion system. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the system architecture of the present invention.

[0019] Figure 2 This is a schematic diagram of the method flow of the present invention.

[0020] Figure 3 This is a schematic diagram of the digital twin model structure of the present invention.

[0021] Figure 4 This is a schematic diagram of the network structure of the deep reinforcement learning decision model of the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figures 1 to 4 The present invention relates to an intelligent processing system and method for energy conversion.

[0024] 1.1 System Architecture Implementation Example Reference Figure 1 The intelligent processing system for energy conversion of this invention comprises five layers: The physical conversion layer comprises multiple energy conversion units. In this embodiment, a natural gas pressure energy power generation system is used as an example, including a turbine expander (rated power 500kW), a generator (rated power 550kW), a heat exchanger, valve assemblies, and other equipment. Each device is equipped with a sensor group: a pressure sensor (range 0-10MPa, accuracy 0.5%FS), a temperature sensor (PT100, accuracy ±0.2℃), a vibration sensor (frequency response 10-1000Hz), a flow meter (accuracy class 1.0), and voltage / current transformers (accuracy class 0.5). Actuators include electric regulating valves (response time <2s) and frequency converters (0-50Hz adjustable).

[0025] The edge computing nodes employ industrial-grade embedded controllers (1.8GHz quad-core CPU, 4GB RAM), deployed within a field control cabinet. They communicate with the sensor array via Modbus TCP / IP protocol, with a sampling frequency set to 100Hz. The edge nodes internally incorporate a sliding window averaging algorithm and a cross-correlation delay estimation algorithm, used to perform preprocessing and feature extraction in step S2.

[0026] The digital twin server uses a high-performance workstation (16 CPU cores, 64GB memory, Tesla T4 GPU) and connects to edge computing nodes via gigabit Ethernet. The server internally deploys a mechanics model (built based on thermodynamic and fluid dynamics equations) and an LSTM error correction model (5 layers, 128 hidden nodes) to perform the simulation and prediction in step S3.

[0027] The central decision server also uses a high-performance workstation configuration and is connected to the digital twin server. The server is equipped with a deep deterministic policy gradient algorithm model, with both the policy network and the value network being three-layer fully connected neural networks (with 256-128-64 nodes respectively), used to execute the decision in step S4.

[0028] The collaborative controller uses a PLC (Programmable Logic Controller), which connects to the actuator via a Profibus DP bus, receives instructions from the central decision server, and converts them into 4-20mA analog or digital signals for transmission.

[0029] The knowledge graph database uses the graph database Neo4j, which stores historical operating condition nodes (approximately 100,000), equipment nodes, fault nodes, policy nodes, and relational edges such as "caused", "solved", and "similar to".

[0030] 1.2 Data Preprocessing and Feature Extraction Examples Reference Figure 2 Step S2, this embodiment details the specific implementation of data preprocessing and feature extraction.

[0031] 1.2.1 Sliding Window Average Method After receiving the raw sensor data, the edge computing node first performs smoothing filtering using the sliding window averaging method. For the voltage data U_i at the i-th sampling point, the filtered value U'_i is calculated using the following formula: U'i = (U{i-4} + U_{i-3} + U_{i-2} + U_{i-1} + U_i + U_{i+1} + U_{i+2}+ U_{i+3} + U_{i+4} + U_{i+5}) / 10 The selection of window size N=10 is based on the following: a sampling frequency of 100Hz corresponds to a sampling interval of 10ms, and 10 points cover a 100ms time window, which can effectively filter out 50Hz power frequency noise and its harmonics while preserving the dynamic response characteristics of the system (the system's dominant time constant is approximately 500ms). Testing showed that using this window improved the signal-to-noise ratio by approximately 15dB.

[0032] 1.2.2 Cross-correlation time delay estimation method Because different sensors have different sampling frequencies and response times (e.g., pressure sensor response time is 5ms, temperature sensor response time is 50ms), it is necessary to align the timestamps of multi-source data. Taking pressure signal x(t) and temperature signal y(t) as an example, the cross-correlation function is calculated: R_{xy}(τ) = ∫ x(t) y(t+τ) dt In the discrete domain implementation, τ is taken from -50ms to +50ms with a step size of 1ms, and the correlation coefficient is calculated for each time delay. The τ value that maximizes R_{xy}(τ) is taken as the time delay estimate τ, and then the temperature signal is shifted by -τ to achieve alignment with the pressure signal. In this embodiment, the measured time delay between pressure and temperature is approximately 35ms.

[0033] 1.2.3 Feature Extraction Three core features were extracted from the aligned data: (1) Equipment health characteristic H_t: generated by comparing the vibration amplitude with the preset alarm threshold. The vibration sensor collects the effective value of the vibration velocity V_rms (unit mm / s) of the turbine expander bearing, and the preset alarm threshold V_max = 4.5 mm / s (according to ISO 10816-3 standard). The health status is calculated as follows: H_t = max(0, 1 - V_rms / V_max) An alarm is triggered when V_rms > V_max, and H_t = 0. In this embodiment, the normal operating condition is V_rms ≈ 2.8 mm / s, corresponding to H_t ≈ 0.38.

[0034] (2) Energy output efficiency characteristic E_t: generated by the ratio of output electrical power to input primary energy power. Output electrical power P_out is obtained by measurement and calculation using voltage and current transformers (P_out = √3 × U × I × cosφ). Input primary energy power P_in is the pressure energy of natural gas, calculated using the following formula: P_in = Q × (p_in - p_out) × η_exp In the formula, Q is the volumetric flow rate (m³ / s). 3 / s), p_in is the inlet pressure (Pa), p_out is the outlet pressure (Pa), and η_exp is the isentropic efficiency of the expander (taken as 0.85). In this embodiment, the typical operating conditions are P_out=425kW, P_in=625kW, and E_t=0.68.

[0035] (3) Environmental coupling characteristic C_t: generated by the difference between ambient temperature and standard temperature. Ambient temperature T_env is collected by an outdoor temperature sensor, and standard temperature T_std is taken as the design operating temperature of 25℃. C_t = T_env - T_std. In this embodiment, the typical summer operating condition is T_env = 35℃ and C_t = +10℃.

[0036] The above three features constitute the state feature set X_t = [H_t, E_t, C_t]^T. A set is generated every 100ms and uploaded to the digital twin server via the MQTT protocol.

[0037] 1.3 Digital Twin Prediction Implementation Examples Reference Figure 2 Step S3 and Figure 3 This embodiment details the specific implementation of the digital twin model.

[0038] 1.3.1 Mechanism Model Construction The mechanistic model is constructed based on the first law of thermodynamics and the governing equations of fluid mechanics. For the turboexpander, the steady-state energy balance equation is adopted: h_in - h_out = W_s + Q_loss In the formula, h_in is the inlet specific enthalpy (kJ / kg), h_out is the outlet specific enthalpy, W_s is the shaft work output per unit mass of working fluid (kJ / kg), and Q_loss is the heat loss (taken as 2% of the input power). The thermophysical properties of the working fluid (natural gas) are calculated using the Peng-Robinson equation of state.

[0039] For generators, the power balance equation is used: P_out = W_s × m × η_gen In the formula, m is the mass flow rate (kg / s), and η_gen is the generator efficiency (taken as 0.96).

[0040] The mechanistic model is solved using the finite difference method, with 100 spatially discretized nodes and a time step of 0.1s. The output is the predicted value within the next 5 seconds, with a time resolution of 0.5s.

[0041] 1.3.2 LSTM Error Correction Model The LSTM error correction model is used to learn the systematic bias of the mechanistic model. The model structure is a 5-layer LSTM, with the following layer configurations: Input layer: Receives the error vectors e_{t-4}, e_{t-3}, ..., e_t from the last 5 time steps. Each error vector has a dimension of 3 (corresponding to the prediction errors of the three features H, E, and C), so the input dimension is 5×3=15.

[0042] LSTM layers 1-3: Each layer has 128 hidden units, using the tanh activation function, and returns a sequence.

[0043] The 4th layer LSTM has 128 hidden units and does not return the sequence (only outputs the last time step).

[0044] Layer 5 Fully Connected: Output dimension 3, corresponding to the prediction error correction values ​​of H, E, and C at the current time.

[0045] The model uses mean squared error (MSE) as the loss function, Adam as the optimizer, a learning rate of 0.001, and a batch size of 32. Training data was collected from historical running records, totaling six months of data, and was divided into training and validation sets in an 8:2 ratio. After 200 epochs of training, the MSE on the validation set converged to 0.0032.

[0046] 1.3.3 Hybrid Prediction Output The final prediction result of the digital twin model is the sum of the output of the mechanistic model and the output of the LSTM error correction model: P_{t+τ} = F_phy(X_t, τ) + G_LSTM(e_{t-4}, e_{t-3}, ..., e_t) In the formula, F_phy represents the output of the mechanistic model, and G_LSTM represents the correction value of the LSTM model output. Calculations are performed for τ = 1s, 2s, ..., 5s to obtain the predicted state sequence within the next 5 seconds, with a time resolution of 1 second. In this embodiment, the mean absolute percentage error (MAPE) of the hybrid model is 2.3%, a significant improvement compared to the pure mechanistic model (MAPE 5.8%).

[0047] 1.4 Deep Reinforcement Learning Decision-Making Examples Reference Figure 2 Step S4 and Figure 4 This embodiment details the specific implementation of the deep reinforcement learning decision model.

[0048] 1.4.1 Definition of State Space and Action Space The state space s_t is composed of the current state feature set X_t (3-dimensional) and the predicted state sequence for the next 5 seconds (5 time points × 3 dimensions = 15 dimensions), for a total of 18 dimensions.

[0049] The action space a_t is a continuous three-dimensional vector: valve opening adjustment Δv (unit: %, range: [-10, 10]), generator speed adjustment Δr (unit: rpm, range: [-100, 100]), and power factor adjustment Δcosφ (range: [-0.1, 0.1]). Each action is a continuous value and is directly output by the policy network.

[0050] 1.4.2 Reward Function Design The reward function R is defined as a weighted sum of three terms: R = 0.5 × η + 0.3 × S - 0.2 × C The calculation methods for each item are as follows: (1) Overall energy conversion efficiency η: η = (P_out + Q_th) / P_in, where P_out is the output electrical power, Q_th is the recovered heat power (recovering turbine exhaust waste heat through a heat exchanger), and P_in is the input power. In this embodiment, Q_th≈150kW, P_out≈425kW, P_in≈625kW, and η≈0.92.

[0051] (2) Safety margin index S: defined as the reciprocal of the degree to which the equipment operating parameters are close to the safety threshold. Taking vibration as an example, S_vib = 1 - (V_rms / V_max)^2, and S_vib approaches 0 when V_rms is close to V_max. The minimum value of multiple parameters (vibration, temperature, pressure) is taken as S. In this embodiment, under normal operating conditions, S≈0.65.

[0052] (3) Operating cost index C: includes equipment wear cost C_wear and energy consumption cost C_energy. C_wear is related to the valve actuation range and speed change rate, while C_energy is related to the power consumption of auxiliary equipment. In this embodiment, C≈0.15.

[0053] 1.4.3 DDPG Network Structure Policy network (Actor) structure: Input layer: 18-dimensional state vector; Hidden layer 1: 256 neurons, ReLU activated, with batch normalization; Hidden layer 2: 128 neurons, ReLU activated; Hidden layer 3: 64 neurons, ReLU activated; Output layer: 3 neurons, tanh activation, output range [-1,1], then linearly mapped to the actual action range; Value network (Critic) structure: Input layer: 18-dimensional state vector + 3-dimensional action vector (concatenated to 21 dimensions); Hidden layer 1: 256 neurons, ReLU activated; Hidden layer 2: 128 neurons, ReLU activated; Hidden layer 3: 64 neurons, ReLU activated; Output layer: 1 neuron, linearly activated, outputs Q-value.

[0054] 1.4.4 Training Process The DDPG algorithm uses the following hyperparameters: Discount factor γ = 0.99; The soft update coefficient τ = 0.001; The experience replay pool has a capacity of 10^6. Batch size 64; Actor learning rate 10^-4, Critic learning rate 10^-3 (Adam optimizer). Noise exploration: Ornstein-Uhlenbeck process, θ=0.15, σ=0.2; Training was conducted offline in a digital twin environment, with a total of 5000 training rounds, each round simulating one hour of operation. After training, the Actor network parameters were fixed and deployed to a central decision server for online decision-making.

[0055] During online operation, a decision is made every 100ms, and the optimal action a_t* is output and executed by the collaborative controller.

[0056] 1.5 Knowledge Graph-Enhanced Decision Making Examples This embodiment details the specific implementation of knowledge graph construction and decision enhancement.

[0057] 1.5.1 Knowledge Graph Construction An energy conversion knowledge graph was constructed using the Neo4j graph database, containing the following node types: Equipment nodes: such as “turbine expander #1”, “generator #2”, “valve #3”, etc., with attributes including equipment model, rated parameters, and installation location.

[0058] Operating condition nodes: such as "inlet pressure 4.5MPa / temperature 15℃ / valve opening 60%", etc., and the attributes include the specific values ​​of each parameter and timestamps.

[0059] Fault nodes: such as "excessive vibration", "excessive temperature", "decrease in efficiency", etc., with attributes including fault code, occurrence time, and handling records. Strategy nodes: such as "increase valve opening by 5%", "reduce load to 80%", "emergency shutdown", etc., with attributes including strategy type and execution effect. Relationship edge types include: "Experience": From the device node to the operating condition node (the operating condition that the device experiences at a certain moment). "Cause": From the working condition node to the fault node (a certain working condition may lead to a certain fault). "Resolve": Policy node, to fault node (a certain policy can resolve a certain fault); "Similar to": Working condition node, to working condition node (the similarity relationship between two working conditions, based on Euclidean distance); A total of 100,000 historical operating condition records, 5,000 fault records, and 2,000 strategy records were imported, and 120,000 nodes and 350,000 relationship edges were constructed.

[0060] 1.5.2 Similarity Matching During online execution, the current state feature set X_t (a 3D vector) is matched with historical work status nodes in the knowledge graph based on similarity. Each historical work status node is pre-stored as a 3D vector (corresponding to health, efficiency, and environmental coupling features). Cosine similarity is used for calculation. sim(X_t, X_his) = (X_t · X_his) / (||X_t|| × ||X_his||) A similarity threshold of 0.85 is set. If a historical work condition with a similarity exceeding the threshold is retrieved, the optimal strategy associated with that work condition is extracted (connected via a "resolution" relationship). In this embodiment, a typical work condition matching result is shown below: Current operating condition: [H=0.38, E=0.68, C=+10] Historical operating condition 1: [H=0.37, E=0.69, C=+9], similarity 0.96, association strategy "valve opening +4.8%"; Historical operating condition 2: [H=0.35, E=0.66, C=+11], similarity 0.91, association strategy "valve opening +5.2%"; Historical operating condition 3: [H=0.30, E=0.65, C=+8], similarity 0.82, below the threshold, not adopted.

[0061] 1.5.3 Strategy Fusion There are two ways to fuse the retrieved historical best strategies: (1) As an initial policy: The retrieved policy action (such as valve opening +4.8%) is used as the initial output of the DDPG policy network, and fine-tuning is performed on this basis. This is equivalent to giving reinforcement learning an "experience starting point" and accelerating convergence.

[0062] (2) As an action constraint: The retrieved strategy is used as the action constraint for the current decision, for example, the range of action variation is limited to not deviating from historical experience by more than ±20%. This embodiment adopts the first method. According to actual tests, after introducing the knowledge graph, the decision convergence time is shortened from an average of 50 steps to 15 steps.

[0063] 1.6 Exception Handling Examples This embodiment details the specific implementation of exception handling.

[0064] 1.6.1 Parameter Exceedance Monitoring Deploy real-time monitoring threads on edge computing nodes to perform threshold determination on sensor data: Temperature monitoring: The temperature alarm threshold for the turbine expander bearing is 100℃. When the detected temperature exceeds 100℃, an emergency shutdown procedure is immediately triggered. The temperature sensor collects data every 10ms, responding via hardware interrupt to ensure a response time of <20ms.

[0065] Pressure monitoring: The inlet pressure alarm threshold is 1.5MPa (1.2 times the design pressure). When the pressure is detected to be >1.5MPa, an emergency shutdown is triggered immediately.

[0066] 1.6.2 Emergency Shutdown Procedure After an emergency stop is triggered, the coordination controller executes the following sequence: Immediately close the inlet shut-off valve (response time <50ms) to stop energy input; Open the bypass pressure relief valve to reduce the system pressure to atmospheric pressure; The generator is unloaded and excitation is stopped; Record abnormal events to the local log and upload them to the monitoring center through redundant communication channels; Issue an audible and visual alarm to notify on-site operators; The entire emergency shutdown process was completed within 300ms, ensuring equipment safety.

[0067] 1.6.3 Communication Interruption Handling The edge computing node communicates with the digital twin server via Ethernet, using a heartbeat mechanism to monitor the connection status with a 1-second heartbeat interval. If no response is received from the server for 5 consecutive seconds, the communication is considered interrupted.

[0068] After communication is interrupted, the edge node automatically switches to local backup control mode. The local backup controller uses a PID algorithm and operates according to a preset maintenance strategy. The valve opening remains unchanged from the last command value before the interruption. If the pressure exceeds 1.2 MPa, close the valve slightly at a fixed slope (-2% / s); If the vibration exceeds 4.0 mm / s, reduce the rotation speed by a fixed slope (-1% / s); Local backup control can maintain system operation safely for at least 30 minutes, pending communication restoration. Once communication is restored, edge nodes automatically synchronize the data from the outage and switch back to centralized decision-making mode.

[0069] 1.7 Performance Indicator Verification Example This embodiment verifies the system performance indicators through actual testing.

[0070] 1.7.1 Response Time Test A test environment was deployed at a natural gas pressure energy power generation site, and a high-precision oscilloscope (1MHz sampling rate) was used to monitor the entire process from sensor data acquisition to command issuance. 100 tests were conducted, and the results are as follows: Minimum response time: 78ms; Maximum response time: 112ms; Average response time: 86ms; 90th percentile response time: 94ms; It meets the design requirement of <100ms.

[0071] 1.7.2 Efficiency Improvement Verification Comparative test plan: Group A (intelligent control of this system) and Group B (traditional PID control) each ran continuously for 7 days, recording input and output power. Test results: Group A's average efficiency: 71.3%; Group B's average efficiency: 65.8%; Efficiency improvement: 5.5%; The t-test showed that p < 0.01, indicating that the difference was statistically significant.

[0072] 1.7.3 Verification of Fault Early Warning Accuracy Historical fault data (including excessive vibration, overheating, and efficiency loss) was collected 120 times, and the data was played back to test the system's fault early warning capability. Test results: Number of correct warnings: 111; Number of missed reports: 9; Number of false alarms: 7; Early warning accuracy rate: 111 / (111+7) = 94.1%; Recall rate: 111 / (111+9) = 92.5%; It meets >90% of the design requirements.

[0073] 1.8 Alternative Embodiments 1.8.1 Application in LNG cold power generation scenarios This system can also be applied to cold energy power generation at LNG receiving terminals. The main adjustments are as follows: Energy conversion unit: adopts a low-temperature turbine expander, with propane / R245fa organic working fluid as the working fluid; Sensor group: Add a low-temperature temperature sensor (-160℃ to +50℃); Mechanism model: Add a two-phase flow heat transfer model to consider the cascade utilization of cold energy; Reward function: Add a cold energy recovery efficiency term, with its weight adjusted to 0.6; 1.8.2 Using the PPO algorithm instead of DDPG Deep reinforcement learning decision-making models can use the Proximal Policy Optimization (PPO) algorithm instead of DDPG. PPO is suitable for discrete action spaces, and its hyperparameters are set as follows: The trimming parameter ε = 0.2; The dominance estimation parameter GAE is λ=0.95; The policy network and the value network share a hidden layer; The learning rate is 3 × 10^-4; Tests showed that the PPO algorithm has slightly better convergence stability than DDPG, but its decision frequency is slightly lower (about 200ms).

[0074] 1.8.3 LSTM replaced with converter The error correction model can use a transformer architecture instead of LSTM. A small-scale transformer (4 layers, 8 attention heads, 256-dimensional hidden layers) is used, with the input being the error sequence of the most recent 10 time steps. Testing shows that the transformer model has a slight advantage in modeling long-term temporal dependencies, but the computational cost increases by approximately 3 times.

[0075] 1.9 Data Structure Definition 1.9.1 Running Data Format (JSON) { "timestamp": 1645564800, / / Unix timestamp, uint64 "device_id": "EXP-001", / / Device ID, string "measurements": { "voltage": 220.5, / / Voltage, float, unit V "current": 10.2, / / Current, float, unit: A "temperature": 85.3, / / Temperature, float, unit °C "pressure": 0.8, / / Pressure, float, unit MPa "flow": 120.5, / / Flow rate, float, unit m 3 / h "vibration": 2.8 / / Vibration, float, unit mm / s } } 1.9.2 State Feature Set Format (Array) [0.38, 0.68, 10.0] / / [Health H, Efficiency E, Environmental Coupling C] 1.9.3 Control Instruction Format (Byte Stream) 1.10 Basis for Parameter Determination 1.10.1 Determining the sliding window size N=10 Based on the sampling theorem and system dynamic characteristics: the system's dominant time constant is approximately 500ms, and a sampling frequency of 100Hz corresponds to a sampling interval of 10ms. The window should cover the main frequency band of the system's dynamic response while avoiding excessive smoothing that could lead to response lag. Spectrum analysis shows that noise is mainly distributed at 50Hz and above; the 10-point window corresponds to a cutoff frequency of approximately 5Hz, effectively filtering out power frequency interference. Field test comparisons show that noise suppression is insufficient when N=5 (signal-to-noise ratio improves by 8dB), response delay increases by 40ms when N=20, and N=10 represents the optimal compromise.

[0076] 1.10.2 Determination of weighting coefficients α=0.5, β=0.3, γ=0.2 The analytic hierarchy process (AHP) was used to determine the weights. Five energy experts were invited to conduct pairwise comparisons of the three indicators—efficiency, safety, and cost—to construct a judgment matrix. The maximum eigenvalue λ_max is calculated to be 3.009, CI = 0.0045, and CR = 0.0087 < 0.1, indicating acceptable consistency. The normalized eigenvectors yield α = 0.54, β = 0.30, and γ = 0.16, which are rounded to α = 0.5, β = 0.3, and γ = 0.2.

[0077] 1.10.3 Basis for the temperature threshold of 100℃ and the pressure threshold of 1.5MPa According to the safety operation manual provided by the equipment manufacturer: The maximum allowable temperature for turbine expander bearings is 95℃; using a safety factor of 1.05, we get 100℃. The design inlet pressure is 1.25 MPa, and with a safety factor of 1.2, it equals 1.5 MPa. The vibration alarm threshold is 4.5 mm / s. According to ISO 10816-3 standard, for rigidly supported rotating machinery, the upper limit of zone C is 4.5 mm / s. 1.10.4 Determining the Number of LSTM Layers and Nodes Optimization was performed using a grid search: combinations of test layers [2, 3, 4, 5, 6] and nodes [32, 64, 128, 256], with validation set MSE as the metric. Results showed that 5 layers and 128 nodes achieved the best balance between model complexity and prediction accuracy (MSE = 0.0032). Increasing the number of layers to 6 only decreased the MSE by 0.0001 but increased the training time by 35%, therefore 5 layers and 128 nodes were chosen.

[0078] 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 smart processing method for energy conversion, characterized in that, Includes the following steps: S1: By deploying sensor groups in each energy conversion unit, operating data is collected in real time at a preset fixed sampling frequency. The operating data includes at least voltage, current, temperature, pressure, and flow rate. S2: The operational data is preprocessed and features are extracted using edge computing nodes to generate a set of state features that characterize the current system operating condition; wherein... The preprocessing includes: using a sliding window averaging method to smooth and filter the data to eliminate random noise, and using a cross-correlation delay estimation method to align the timestamps of sensor data with different sampling frequencies; The feature extraction includes: generating equipment health features based on the comparison between vibration amplitude and preset alarm threshold; generating energy output efficiency features based on the ratio of output power to input primary energy power; and generating environmental coupling features based on the difference between ambient temperature and standard temperature. S3: Input the state feature set into the digital twin model, which contains two parallel computing sub-modules: The first submodule is a mechanism model built based on thermodynamic and fluid dynamic equations, used to simulate the physical operating state of the energy conversion unit according to the current state feature set; The second submodule is an error correction model built on a long short-term memory neural network, which is used to predict the output error of the mechanism model at the current moment by taking the prediction errors of the most recent multiple time points as input. The simulation results of the mechanism model are added to the prediction error output by the error correction model to obtain the corrected prediction result, and the predicted state sequence within the future preset time period is output based on this. S4: Input the state feature set and predicted state sequence into a decision model based on deep reinforcement learning. The decision model adopts a deep deterministic policy gradient algorithm, using a multi-objective function containing energy conversion efficiency, safety margin, and operating cost as the reward function. Through the interactive iteration of the policy network and the value network, the optimal control action that maximizes the cumulative reward is solved. S5: Convert the optimal control action into control commands and send them to the actuators of each energy conversion unit to adjust the valve opening, generator speed or power factor; S6: Repeat steps S1 to S5 to form a closed-loop control.

2. The processing method according to claim 1, characterized in that: In S2, the sliding window averaging method is used for smoothing filtering with a window size of N=10. The cross-correlation delay estimation method calculates the correlation coefficient of the two signals under different delays and takes the delay value when the correlation coefficient is the largest as the alignment basis.

3. The processing method according to claim 1, characterized in that: The long short-term memory neural network in S3 has a 5-layer structure with 128 hidden layer nodes. The input is the error vector between the mechanism model output and the actual value in the last 5 time steps, and the output is the prediction error value at the current time step.

4. The processing method according to claim 1, characterized in that: The reward function of the deep deterministic policy gradient algorithm in S4 is as follows: in, To improve overall energy conversion efficiency, As a safety margin indicator, As an operating cost indicator, , , The preset weight coefficients are used; the strategy network is a 3-layer fully connected neural network, and the value network is a 3-layer fully connected neural network.

5. The processing method according to claim 1, characterized in that: It also includes exception handling steps: After step S1, the operating data is monitored in real time to see if it exceeds the preset safety threshold. If the temperature exceeds 100°C or the pressure exceeds 1.5MPa, an emergency shutdown is triggered immediately. If communication is interrupted for more than 5 seconds, the local backup controller will be activated and operation will continue according to the preset maintenance strategy.

6. The processing method according to claim 1, characterized in that: The method also includes a knowledge graph-based decision enhancement step: Before step S4, the current state feature set is matched with the pre-constructed energy conversion knowledge graph for similarity matching. Historical operating conditions and their optimal control strategies with a cosine similarity greater than 0.85 are retrieved, and the retrieval results are used as the initial policy or action output constraints of the policy network in the decision model.

7. The processing method according to claim 1, characterized in that: The performance metrics of the method include: The system response time is less than 100 milliseconds, from data acquisition to command issuance; Energy conversion efficiency is improved by more than 5% compared to a baseline system that does not employ intelligent control; The accuracy rate of equipment fault early warning exceeds 90%, which is verified based on historical fault data.

8. An intelligent processing system for energy conversion, used to execute the processing method according to any one of claims 1-7, characterized in that, include: The physical conversion layer contains multiple energy conversion units, each equipped with a sensor array and an actuator; Edge computing nodes, which are communicatively connected to the sensor group, are used to perform preprocessing and feature extraction in step S2; A digital twin server, connected to an edge computing node, is used to perform the simulation and prediction in step S3; The central decision server, connected to the digital twin server, is internally deployed with a deep reinforcement learning model for executing the decision in step S4; The collaborative controller is connected to both the central decision-making server and the execution mechanism, and is used to issue instructions for step S5.

9. The processing system according to claim 8, characterized in that, The central decision server is also connected to a knowledge graph database, which stores historical operating conditions and corresponding optimal control strategies for decision enhancement.

10. A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the processing method according to any one of claims 1-7.

Images