Shale oil station temperature closed loop prediction method

Abstract

The present application relates to the technical fields of automation control and industrial signal processing, and discloses a shale oil station temperature closed-loop prediction method, comprising: obtaining current sampling of outlet temperature and flow data; online adjusting variational mode decomposition scale according to dynamic residual error energy entropy index, decoupling temperature sampling data into determined main trend component and multi-scale time-frequency disturbance component; inputting each component into a multivariate time series cascade prediction network to calculate initial prediction value; dynamically reconstructing sliding time domain data matrix internal circulating queue storage capacity upper limit according to flow data to in-situ align time phase; correcting initial prediction value by a mutation variable isolation gate unit and outputting target predicted temperature; comparing process temperature control target value to output control valve action increment instruction; the present application realizes in-situ alignment of feature matrix and heat transfer phase by reconstructing sliding time domain window scale, and eliminates trend prediction overshoot of variable flow large lag system.

Description

Technical Field

[0001] This invention relates to a closed-loop prediction method for temperature at shale oil fields, belonging to the field of automation control and industrial signal processing technology. Background Technology

[0002] Currently, in fluid heating control systems, temperature closed-loop control for controlled objects with large time lag generally adopts an adaptive trend prediction control strategy based on historical sampling sequences. This strategy involves collecting discrete fluid temperature signals from sensors and constructing a time-domain feature matrix to predict temperature evolution characteristics over multiple future steps. Based on this, heating source adjustment parameters are calculated to stabilize the outlet temperature. However, hardware improvements such as furnace structure modifications and optimized pipeline insulation are insufficient to overcome the bottleneck of large heat transfer lag. Software control algorithms remain the key path. For example, Chinese invention patent application CN107621781A discloses a method and system for constant-temperature heating and self-optimizing thermal efficiency control of an oilfield furnace. This method utilizes a fuzzy neural network to adjust fuel flow based on instantaneous temperature deviation and rate of change. However, this method implicitly relies on the underlying physical property that the number of heat transfer lag steps is relatively constant. In the face of variable flow rate conditions at shale oil fields, dynamic flow fluctuations cause nonlinear slippage in the heat transfer lag span, leading to a mismatch between the existing fixed sampling architecture and the actual heat transfer phase. The feature matrix cannot adaptively align with the time phase, causing control loop overshoot and making it difficult to guarantee temperature control stability under variable conditions.

[0003] In the gathering and transportation of fluids with high wax content and high freezing point, the temperature control loop must maintain a stable thermal temperature field. Fluid heat conduction exhibits time lag characteristics, and the heat transfer lag is physically coupled with the volumetric flow rate. When the feed flow rate fluctuates dynamically, the heat transfer lag span undergoes nonlinear slippage. Conventional strategies employ a fixed sampling period and a constant time-domain window size. However, under variable flow rate and pressure conditions, the pure lag time shortens with increasing flow rate and lengthens with decreasing flow rate. The fixed sampling architecture causes a misalignment between the window time-domain reference and the actual heat transfer phase. This phase misalignment results in characteristic ambiguity in the data matrix of the input prediction loop. The inability to separate high-frequency random disturbances from mid-to-low-frequency main trend components leads to overshoot oscillations in the control loop, causing frequent alternating fatigue in the fuel supply valve. To mitigate time-delay disturbances caused by flow rate variations, conventional approaches typically employ conservative strategies such as directly reducing the controller gain or extending the control waiting time. While this approach can alleviate overshoot, it reduces response sensitivity and cannot handle sudden large-scale disturbances. Another attempt was to introduce a filter at the input to smooth signal fluctuations, but the static filtering mechanism introduces group delay, resulting in a superposition of secondary phase lags in the control loop, exacerbating the mismatch between data characteristics and the actual heat transfer phase.

[0004] Therefore, the technical problem to be solved by this invention is how to dynamically reconstruct the time-domain physical scale of the sliding window based on the real-time volumetric flow rate, so as to achieve in-situ alignment between the discrete feature matrix and the real heat conduction phase while suppressing high-frequency random disturbances and eliminating trend prediction overshoot of variable flow rate large lag system. Summary of the Invention

[0005] To address the problems in the background technology, this invention addresses the issues of predicted phase misalignment and control overshoot oscillation in large-delay control systems under varying operating conditions. The technical solution is as follows: A closed-loop temperature prediction method for shale oilfields, comprising the following steps:

[0006] Step S101: Obtain the current temperature sampling data of the furnace outlet and the fluid volume flow rate data of the furnace feed.

[0007] Step S102: Call the variational mode decomposition model, and adjust the decomposition scale of the variational mode decomposition model online based on the dynamic residual energy entropy index obtained by asynchronous empirical mode decomposition of the current temperature sampling data, so as to decouple the current temperature sampling data into monotonically variable main trend components and multi-scale time-frequency disturbance components to isolate random noise.

[0008] Step S103: Input the main trend component and the multi-scale time-frequency perturbation component into the multivariate time-series cascaded prediction network, use the one-dimensional convolutional layer to extract the local time-frequency feature map, and use the long short-term memory network layer to capture the long-term time evolution trend, and output the initial predicted value of the furnace outlet temperature.

[0009] Step S104: Run the flow adaptive window scale update program, calculate the time-domain phase discrepancy number caused by flow velocity changes based on fluid volume flow rate data and inherent volume constant of the pipeline network, reconstruct the upper limit of the storage capacity of the circular queue inside the sliding time-domain data matrix to align the physical time phase of the historical comparison sequence in situ when the flow velocity changes, and call the trend consistency identification operator to apply first-order discrete difference operation to the historical comparison sequence to generate the first symbol vector and the second symbol vector respectively. Generate the trend consistency index through point-by-point same sign verification. The mutation isolation gating unit corrects the initial prediction value according to the mapping relationship between the trend consistency index and the transient prediction error to output the target prediction temperature. Use a digital comparator to perform difference comparison operation between the target prediction temperature and the process temperature control target value to output the control valve action increment command, and send it to the electro-hydraulic drive device of the fuel supply valve to adjust the valve opening.

[0010] Preferably, step S102 includes the following sub-steps: Step S1021, performing empirical mode decomposition on the current temperature sampling data, and calculating the envelope spectrum entropy value of the obtained residual sequence as a dynamic residual energy entropy index; Step S1022, if the difference between the envelope spectrum entropy value and the set entropy benchmark value exceeds the set fluctuation threshold, determining that the current temperature sampling data contains non-stationary random noise and triggering the decomposition scale increment signal; Step S1023, increasing the number of mode decompositions of the variational mode decomposition model according to the decomposition scale increment signal until the energy change rate of the residual sequence is lower than the set convergence threshold.

[0011] Preferably, step S103 includes the following sub-steps: Step S1031, the main trend component and the multi-scale time-frequency perturbation component are matrix-concatenated to construct a multivariate parallel input matrix; Step S1032, the multivariate parallel input matrix is ​​fed into a one-dimensional convolutional layer, and a sliding window convolution kernel of a set size is used to perform feature sliding window scanning along the time axis to extract local time-frequency feature maps; Step S1033, the local time-frequency feature maps are flattened and input into a long short-term memory network layer, and the forget gate, input gate and output gate gate in the layer are used to update and decay the historical state to capture the long-term time evolution trend and output the initial prediction value.

[0012] Preferably, the step S104, which involves running the flow adaptive window scale update program and aligning the physical time phase of the historical comparison sequence in situ, includes the following sub-steps: Step S1041, reading fluid volumetric flow rate data from the data bus in real time, and calculating the corresponding time-domain delay deviator number under the current flow rate condition by calling the geometric transfer function based on the inherent volume constant of the pipeline network; Step S1042, reconstructing the upper limit of the storage capacity of the circular queue inside the sliding time-domain data matrix based on the time-domain delay deviator number, and dynamically increasing or decreasing the storage length of the circular queue to align the physical time phase of the historical comparison sequence and the current temperature sampling data in situ when the flow rate changes.

[0013] Preferably, the step S104, which involves calling the trend consistency identification operator to generate a trend consistency index through point-by-point same sign verification, includes the following sub-steps: Step S1043, applying first-order discrete difference operation to the historical comparison sequence to generate a first symbol vector and a second symbol vector respectively; Step S1044, performing point-by-point same sign verification on the first symbol vector and the second symbol vector and performing time-domain point multiplication and accumulation operation within a set time sliding window to generate a trend consistency index.

[0014] Preferably, step S104, which uses a digital comparator to perform a difference comparison calculation between the target predicted temperature and the process temperature control target value to output an incremental control valve action command, includes the following sub-steps: Step S1045, inputting the target predicted temperature to the first input terminal of the digital comparator, and inputting the process temperature control target value read from the main control memory to the second input terminal of the digital comparator; Step S1046, using the digital comparator to perform a difference comparison calculation between the target predicted temperature and the process temperature control target value, outputting the corresponding incremental control valve action command, and sending it in real time to the electro-hydraulic drive device of the fuel supply valve through the digital output port.

[0015] Preferably, step S101 includes the following sub-steps: step S1011, collecting real-time temperature data of the furnace outlet to generate current temperature sampling data; step S1012, synchronously reading the real-time volumetric flow rate of the furnace feed through the flow meter data bus to generate fluid volumetric flow rate data.

[0016] Preferably, after step S104, the following step is also included: Step S105, in each closed-loop control cycle, the target predicted temperature is compared with the set safety upper limit temperature by amplitude. If the target predicted temperature is greater than the safety upper limit temperature, an emergency fuel cut-off command is triggered and the control valve action increment command is forcibly rewritten as a safety valve closing command to close the fuel supply valve.

[0017] Preferably, the entire prediction and control process runs spontaneously in a closed loop with a fixed control cycle. At the beginning of each control cycle, the latest current temperature sampling data is reacquired and the corresponding control valve action increment command is recalculated. After the material flow fluctuation triggers adaptive feedback adjustment, the furnace outlet temperature is stabilized within the error control range set by the process.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] 1. In the closed-loop prediction of temperature at shale oil fields, the original historical sequence of outlet temperature is input to the variational mode decomposition module. The variational constraint decomposition operator is called to decouple the temperature into multiple intrinsic sub-mode components and trend residuals. The central processing unit monitors the sample energy entropy value of the previous prediction output sequence in real time. When the sample energy entropy value exceeds the preset critical health threshold, the main control chip automatically increases the secondary penalty factor in the variational constraint decomposition operator to achieve adaptive attenuation of high-frequency step interference, block the penetration of undecomposed high-frequency noise into the mid- and low-frequency main trend components, and enable the data matrix input to the prediction loop to have monotonic deterministic characteristics under complex variable operating conditions, thereby improving the time-frequency analysis accuracy of the input data of the closed-loop control loop.

[0020] 2. The standardized components are input in parallel to the multivariate time-series cascaded prediction network. The convolutional layer uses a sliding convolutional kernel with a set stride to extract time-frequency feature maps that reflect local high-frequency disturbances or mid-frequency periodic fluctuations in temperature. The long short-term memory network layer uses the forget gate, input gate, and output gate gating mechanism to control the updating and decay of historical states and capture long-term change trends. This cascaded prediction mechanism of local time-frequency features and long-term time series dependence avoids the prediction divergence defects that are prone to occur when a single time-series prediction model processes aliased noise temperature signals, and provides a stable multi-step trend prediction benchmark for controlled objects with large inertia and large hysteresis of deep thermal conduction.

[0021] 3. The feedback correction unit initiates the flow adaptive window scale update procedure. The main control chip reads the current feed volume flow rate from the data bus in real time and calls the geometric transfer function in combination with the internal equivalent fluid volume constant of the internal pipe network to dynamically calculate the discrete step number corresponding to the flow rate change under the current operating condition. This causes the sliding time domain data matrix to dynamically reconstruct the upper limit of the internal circulating queue storage capacity. When the material flow rate changes, the physical time phase of the historical comparison sequence is aligned in situ. This eliminates the trend analysis phase misalignment caused by the flow rate change under the traditional fixed step window under changing operating conditions, and avoids the control loop from misjudging the error attribute under the changing flow rate state. Attached Figure Description

[0022] Figure 1 This is a flowchart of a closed-loop prediction method for the outlet temperature of a shale oil field heater according to the present invention;

[0023] Figure 2 This is a diagram illustrating the architecture of a closed-loop prediction system for the outlet temperature of a shale oil field heater.

[0024] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0026] A closed-loop prediction method for temperature at shale oil fields includes the following steps:

[0027] Step S101: Obtain the current temperature sampling data of the furnace outlet and the fluid volume flow rate data of the furnace feed.

[0028] Step S102: Call the variational mode decomposition model, and adjust the decomposition scale of the variational mode decomposition model online based on the dynamic residual energy entropy index obtained by asynchronous empirical mode decomposition of the current temperature sampling data, so as to decouple the current temperature sampling data into monotonically variable main trend components and multi-scale time-frequency disturbance components to isolate random noise.

[0029] Step S103: Input the main trend component and the multi-scale time-frequency perturbation component into the multivariate time-series cascaded prediction network, use the one-dimensional convolutional layer to extract the local time-frequency feature map, and use the long short-term memory network layer to capture the long-term time evolution trend, and output the initial predicted value of the furnace outlet temperature.

[0030] Step S104: Run the flow adaptive window scale update program, calculate the time-domain phase discrepancy number caused by flow velocity changes based on fluid volume flow rate data and inherent volume constant of the pipeline network, reconstruct the upper limit of the storage capacity of the circular queue inside the sliding time-domain data matrix to align the physical time phase of the historical comparison sequence in situ when the flow velocity changes, and call the trend consistency identification operator to apply first-order discrete difference operation to the historical comparison sequence to generate the first symbol vector and the second symbol vector respectively. Generate the trend consistency index through point-by-point same sign verification. The mutation isolation gating unit corrects the initial prediction value according to the mapping relationship between the trend consistency index and the transient prediction error to output the target prediction temperature. Use a digital comparator to perform difference comparison operation between the target prediction temperature and the process temperature control target value to output the control valve action increment command, and send it to the electro-hydraulic drive device of the fuel supply valve to adjust the valve opening.

[0031] Preferably, step S102 includes the following sub-steps: Step S1021, performing empirical mode decomposition on the current temperature sampling data, and calculating the envelope spectrum entropy value of the obtained residual sequence as a dynamic residual energy entropy index; Step S1022, if the difference between the envelope spectrum entropy value and the set entropy benchmark value exceeds the set fluctuation threshold, determining that the current temperature sampling data contains non-stationary random noise and triggering the decomposition scale increment signal; Step S1023, increasing the number of mode decompositions of the variational mode decomposition model according to the decomposition scale increment signal until the energy change rate of the residual sequence is lower than the set convergence threshold.

[0032] Preferably, step S103 includes the following sub-steps: Step S1031, the main trend component and the multi-scale time-frequency perturbation component are matrix-concatenated to construct a multivariate parallel input matrix; Step S1032, the multivariate parallel input matrix is ​​fed into a one-dimensional convolutional layer, and a sliding window convolution kernel of a set size is used to perform feature sliding window scanning along the time axis to extract local time-frequency feature maps; Step S1033, the local time-frequency feature maps are flattened and input into a long short-term memory network layer, and the forget gate, input gate and output gate gate in the layer are used to update and decay the historical state to capture the long-term time evolution trend and output the initial prediction value.

[0033] Preferably, the step S104, which involves running the flow adaptive window scale update program and aligning the physical time phase of the historical comparison sequence in situ, includes the following sub-steps: Step S1041, reading fluid volumetric flow rate data from the data bus in real time, and calculating the corresponding time-domain delay deviator number under the current flow rate condition by calling the geometric transfer function based on the inherent volume constant of the pipeline network; Step S1042, reconstructing the upper limit of the storage capacity of the circular queue inside the sliding time-domain data matrix based on the time-domain delay deviator number, and dynamically increasing or decreasing the storage length of the circular queue to align the physical time phase of the historical comparison sequence and the current temperature sampling data in situ when the flow rate changes.

[0034] Preferably, the step S104, which involves calling the trend consistency identification operator to generate a trend consistency index through point-by-point same sign verification, includes the following sub-steps: Step S1043, applying first-order discrete difference operation to the historical comparison sequence to generate a first symbol vector and a second symbol vector respectively; Step S1044, performing point-by-point same sign verification on the first symbol vector and the second symbol vector and performing time-domain point multiplication and accumulation operation within a set time sliding window to generate a trend consistency index.

[0035] Preferably, step S104, which uses a digital comparator to perform a difference comparison calculation between the target predicted temperature and the process temperature control target value to output an incremental control valve action command, includes the following sub-steps: Step S1045, inputting the target predicted temperature to the first input terminal of the digital comparator, and inputting the process temperature control target value read from the main control memory to the second input terminal of the digital comparator; Step S1046, using the digital comparator to perform a difference comparison calculation between the target predicted temperature and the process temperature control target value, outputting the corresponding incremental control valve action command, and sending it in real time to the electro-hydraulic drive device of the fuel supply valve through the digital output port.

[0036] Preferably, step S101 includes the following sub-steps: step S1011, collecting real-time temperature data of the furnace outlet to generate current temperature sampling data; step S1012, synchronously reading the real-time volumetric flow rate of the furnace feed through the flow meter data bus to generate fluid volumetric flow rate data.

[0037] Preferably, after step S104, the following step is also included: Step S105, in each closed-loop control cycle, the target predicted temperature is compared with the set safety upper limit temperature by amplitude. If the target predicted temperature is greater than the safety upper limit temperature, an emergency fuel cut-off command is triggered and the control valve action increment command is forcibly rewritten as a safety valve closing command to close the fuel supply valve.

[0038] Preferably, the entire prediction and control process runs spontaneously in a closed loop with a fixed control cycle. At the beginning of each control cycle, the latest current temperature sampling data is reacquired and the corresponding control valve action increment command is recalculated. After the material flow fluctuation triggers adaptive feedback adjustment, the furnace outlet temperature is stabilized within the error control range set by the process.

[0039] Example 1

[0040] In the closed-loop control scenario of the heater outlet temperature at the Changqing shale oil site, the system faces frequent random fluctuations in crude oil feed flow rate, causing the thermal heating system to exhibit large lag and nonlinear time-varying characteristics. The heater outlet temperature sequence is affected by fluctuations in crude oil feed flow rate, inlet temperature, and fuel calorific value, exhibiting non-stationary characteristics. To perform temperature closed-loop prediction, the acquisition unit obtains heater outlet temperature data in real time. and feed volume flow rate data The central processing unit runs the variational mode decomposition module and calls the quadratic penalty factor. and Lagrange multipliers The variational constraint decomposition operator formed will Adaptive decoupling into 5 intrinsic submodal components to and one trend residual term During the actual control cycle, the system adopts a time-sharing asynchronous execution logic and a dual-core parallel hardware architecture. Specifically, the empirical mode decomposition (EMD) process is confined to an independent signal processing coprocessor for offline asynchronous operation, and the data stream window length is fixed and shortened to a local micro-region containing only 50 historical sampling points. By limiting the number of internal loop iterations to no more than 3, the single calculation time of EMD is ensured to be controlled within 2 milliseconds. The envelope spectrum entropy value output by the coprocessor is sent to the main core of the main control chip every 100 milliseconds via a direct memory access channel. The main core triggers online adjustment of the variational mode decomposition scale only when it receives this signal. During the rest of the time, the main core independently and continuously executes the variational mode decomposition of the current control cycle with a fixed decomposition scale, thereby avoiding synchronous calculation of two algorithms on the same time axis to ensure the continuous stability of the main closed-loop control loop. During the prediction cycle of variational mode decomposition, the main control chip calculates the sample energy entropy value of the preceding predicted output sequence. ,when Exceeding the preset critical health threshold At that time, the main control chip automatically activates the parameter scaling correction operator, and... The value is increased step by step according to a preset ratio to enhance the attenuation capability against high-frequency step interference.

[0041] Multivariate temporal cascaded prediction network prediction module processes the normalized components. The normalization calculation uses the formula ,in, For the corresponding original component data, and These represent the maximum and minimum values ​​of the component sequence within the current sliding statistical window. Each component's time-frequency feature map is extracted via a one-dimensional convolutional layer and input into a long short-term memory network layer. Sub-predicted values ​​are then calculated using forget gates, input gates, and output gates. to and residual prediction values In this process, the one-dimensional convolutional layer is configured with 16 one-dimensional sliding window convolutional kernels, each with a temporal dimension of 3 and a sliding stride of 1 along the time axis to achieve lossless local time-frequency feature extraction. The long short-term memory network layer adopts a single-layer bidirectional architecture, with 64 nodes in its hidden layer and a fully connected layer at the output to convert the 64-dimensional hidden state features into a single-dimensional temperature sub-prediction value output. The superimposed reconstruction unit calls the dimensionless inverse mapping operator and performs linear summation to obtain the initial predicted value of the outlet temperature for the current prediction period. The feedback correction unit executes the flow adaptive window scale update procedure. To clarify the specific execution logic of the two operators mentioned above, when the parameter variable scale correction operator is running advanced adjustment, if the sample energy entropy value exceeds the critical health threshold of 0.15, it automatically triggers a step increment instruction, increasing the secondary penalty factor by 500 each time until it reaches the upper limit of 5000. The dimensionless inverse mapping operator performs the inverse operation opposite to normalization, multiplying the dimensionless sub-predicted value between 0 and 1 output by the difference between the maximum and minimum values ​​of the corresponding component in the current sliding statistical window, and superimposing the minimum value to restore it to a physical value of ℃ with the true temperature dimension. Finally, the true temperature predicted values ​​of each component are algebraically summed to output the final initial predicted temperature. The main control chip reads the current cycle from the data bus. The geometric transfer function is invoked in conjunction with the inherent equivalent fluid volume constant of the internal piping network. Through calculation formula The time-domain delay deviance number corresponding to the current flow velocity condition is calculated. ,in, To ensure precise alignment between the overall physical dimensions and the discrete step counts of the control loop, the system's base sampling frequency is introduced for discretization based on the time delay obtained by dividing the equivalent fluid volume constant by the fluid volumetric flow rate. Specifically, the physical unit of the ratio of the equivalent fluid volume constant to the fluid volumetric flow rate is hours. This is multiplied by 3600 seconds to convert it to seconds, and then multiplied by the system's set temperature data acquisition frequency of 10 Hz or 20 Hz. The resulting product is the total number of delay sampling points under the current flow rate condition. After rounding this total number of delay sampling points using a rounding operator, the final time-domain delay discrete step count is output. The sliding time-domain data matrix converges the storage capacity limit of its internal circular queue to... It continuously records the historical queue of actual measured temperature and the historical queue of initial predicted temperature, so as to achieve physical time phase alignment between the historical comparison sequence and the current temperature sampling data.

[0042] The trend consistency identification operator performs a first-order discrete difference operation on the queue, extracting the first sign vector of the measured trajectory trend and the second sign vector of the predicted trajectory trend. Through point-by-point same-sign verification and time-domain dot product accumulation, a trend consistency index ranging from -1 to 1 is generated. Transient dynamic prediction error From the formula The calculation shows that, For actual temperature measurement, The value is the predicted value for the previous period. Less than or equal to the maximum safety boundary threshold At that time, the system maintains a constant smoothing filter gain coefficient. If running, ,when At that time, the system activates the mutation truncation mode, using the formula... Achieve gain coefficient The step decay, where The value ranges from 0.1 to 0.3; when At that time, the system activates phase-tracking mode to maintain... Input the complete value and calculate the delay phase compensation bias. In phase-tracking mode, the method for obtaining the delayed phase compensation bias is as follows: The transient dynamic prediction error is calculated by multiplying it by the current cycle trend consistency index to obtain the trend-weighted error value. This trend-weighted error value is then multiplied by a fixed coefficient of 0.5 to serve as the basic phase correction. Finally, the first-order forward differential sign of the actual measured temperature within the current control cycle is extracted. When the differential value is positive, the sign of the basic phase correction remains unchanged; when the differential value is negative, the basic phase correction is inverted. This generates a directional delayed phase compensation bias, targeting the predicted temperature. The calculation formulas are respectively or The control command generation unit will Compared with the process temperature control target value, the output controls the valve's action increment command, driving the electro-hydraulic drive device of the fuel supply valve to adjust the valve opening. When the temperature exceeds the upper limit of safety, the incremental command for controlling the valve action is rewritten as a safe valve closing command, and the outlet temperature of the heating furnace stabilizes within the error control range set by the process.

[0043] Example 2

[0044] This embodiment aims to verify the accuracy and stability of the invention in predicting the outlet temperature of a shale oil field heater under conditions of frequent fluctuations in crude oil feed flow. An experimental platform was constructed, comprising a simulated heating circuit and a data acquisition and processing terminal. The heater simulation component includes… The rated thermal power and crude oil physical properties of 0.92 g / cm³ were used to simulate the variable flow conditions in the production site by introducing a sinusoidal fluctuation with a period of 0.5 Hz and a random step pulse sequence. The sampling frequency of the system data acquisition unit was set to 10 Hz to ensure that the transient time-domain characteristics of the heater outlet temperature under flow fluctuations could be captured. The experimental comparison sample group design is as follows: the conventional proportional-integral-differential algorithm was used as the first control group, and the formula-based algorithm was removed. The system with dynamic reconstruction logic of the time-domain data matrix was used as a second control group to remove the sample energy entropy value. Decomposition penalty factor The dynamically corrected system serves as the third control group, when the feed volumetric flow rate... When the flow rate was reduced from 12.5 m³ / h to 8.2 m³ / h, the data for each group were recorded as follows: The first control group, due to the lack of effective prediction of hysteresis characteristics, measured a maximum outlet temperature overshoot of 14.8℃, and it took 420s to converge to the target setpoint. In order to accurately verify the independent technical contribution of the flow adaptive window mechanism, a single variable control analysis was performed under the same multivariate time-series cascaded prediction network structure and control gain parameters. The second control group, which had the flow adaptive window matrix dynamic reconstruction logic turned off, had a phase deviation of 85s between the measured temperature and the predicted temperature due to the lag in the time delay queue update, resulting in a predicted value hysteresis deviation of 6.3℃, and directly causing the loop to generate a maximum outlet temperature overshoot of 7.5℃.

[0045] The experimental group of this invention, after enabling only the traffic adaptive window scale update program, calculated the delay steps based on the real-time changes in traffic data using the traffic adaptive window scale update program. The dynamic convergence period was reduced from 120 cycles to 185 cycles, ensuring in-situ alignment of the characteristic matrix with the heat transfer phase of the pipeline network. A phase compensation bias of 1.8℃ was injected online through logical determination. The measured maximum outlet temperature overshoot was significantly reduced from 7.5℃ in the second control group when the mechanism was disabled to 1.2℃, and the fluctuation convergence time was shortened to 115s. Extreme value verification of the numerical boundary showed that when the penalty factor in the variational constraint decomposition operator... When the value is 500, the energy entropy When the value is 0.12, the system performs stably; when Increasing the value to 5000 improves the system's prediction artifact suppression efficiency by 35% under high-frequency step disturbances. As the value is increased to 8000, the computational matrix becomes too rigid, and the predicted data sequence shows a significant decrease in coherence. This indicates that 500 to 5000 is the effective parameter range for the operator under the current computing power constraints. The above data proves that by dynamically reconstructing the characteristic parameters at the data processing end, it is possible to effectively decouple unsteady flow disturbances and thermal inertia hysteresis, and achieve a balance between prediction accuracy and control stability.

[0046] Example 3

[0047] During the operation of the heating furnace in the shale oil field gathering and transportation system, to address the thermal inertia nonlinear hysteresis problem caused by fluctuations in crude oil properties, this embodiment of the invention constructs a predictive control closed loop that includes phase dynamic compensation and abrupt signal gating to eliminate the amplitude overshoot of control commands. The initial system state is defined as follows: the controlled pipeline pressure of the heating furnace is maintained at 1.5 MPa, the density of the feed crude oil is 0.92 g / cm³, the temperature sensor acquisition frequency is set to 20 Hz, the measurement resolution is 0.05℃, and the central processing unit reads the feed volumetric flow rate in real time. Real-time pressure of pipeline network The flow adaptive window scale update program calls the pipeline pressure compensation operator, reconstructing the original delay discrete step count calculation model based on a single flow rate into the following formula: ,in, The system's temperature data acquisition frequency is set to , where is the equivalent fluid volume constant of the pipe network. The fluid compressibility coefficient, The set reference pressure for the pipeline network, As a rounding function, the model corrects the influence of flow velocity on hysteresis through a pressure compensation factor, achieving in-situ alignment of the data matrix and heat transfer phase, eliminating prediction bias caused by pressure fluctuations. The execution procedure of the mutation isolation gating unit includes the following three logical steps: In specific implementation, based on the influence of fluid compressibility characteristics on the time delay discrete step count, the fluid compressibility coefficient is set as the equivalent engineering compensation coefficient, with a value of 0.02 MPa. When the real-time pressure of the pipeline deviates from the reference pressure by 0.5 MPa, the pressure correction coefficient calculated by the above equivalent engineering compensation coefficient will cause a 1% virtual scaling of the flow term. This scaling, through the cumulative effect, can cause the product result to break through the rounding threshold when passing through a large sample sliding window, thereby causing a fine adjustment of 1 to 2 cycles in the final output time-domain time delay discrete step count, realizing the quantitative mapping of surface physical characteristics at the discrete control level, through historical error sequences. Calculate the moving standard deviation within the current sliding window. ,use For the preset maximum safety boundary threshold Perform dynamic scaling to generate dynamic thresholds The calculation formula is: ,in, This is the sensitivity adjustment coefficient. The transient dynamic prediction error is represented by the reference standard deviation. From the formula Calculate, where, For actual temperature measurement, If the value is the predicted value from the previous period, then... Exceed The system reads the transient change state vector, and when the trend consistency index... At that time, the gain coefficient The decay logic is executed using the following formula: ,in, To adjust the attenuation constant, this logic automatically increases the system damping when the error exceeds the limit, preventing the control output from diverging.

[0048] exist In the calibration procedure, the system executes an offline calibration process, recording the root mean square distribution of temperature prediction deviation within 10 seconds after a sudden flow change, selecting 2.5℃ as the baseline. At the set value, 95% of flow step disturbances are effectively isolated, and the misjudgment rate under normal flow velocity fluctuations is 0, at the target predicted temperature. When in a noncoherent trend, the phase compensation bias is invoked. By performing predictive phase pre-adjustment, the final test data showed that under a pipeline pressure of 1.5 MPa, the feed flow rate suddenly dropped from 12.5 m³ / h to 8.2 m³ / h, the maximum overshoot of the heating furnace outlet temperature was reduced to 1.2℃, and the fluctuation convergence period was shortened to 85 s, thus achieving stable prediction and control of the nonlinear hysteresis process.

[0049] Example 4

[0050] During the initial deployment phase of an industrial heating furnace, the system needs to undergo a standardized calibration procedure based on physical characteristics to determine the smoothing filter gain coefficient in the feedback correction unit. The calibration procedure selects the heating furnace operating under rated heat load and without sudden changes in crude oil properties as the reference environment. The system injects a sinusoidal disturbance signal with an amplitude of 2.0℃, records the output response curve of the outlet temperature, and calculates the phase lag angle of the closed-loop control system at a frequency of 0.1Hz. Gain coefficient The calibration basis is: ,in, This represents the maximum permissible filter gain of the system in steady state. The reference phase hysteresis threshold for the heat conduction model of the heating furnace is determined when the measured value is... When the angle is 15°, the calibration program automatically outputs... The initial calibration value is 0.15. This procedure directly anchors the control logic to the geometric volume and thermal conductivity physical characteristics of the heating furnace, ensuring that heating furnaces of different specifications have a quantified gain adjustment benchmark before being put into use, eliminating response hysteresis caused by individual equipment differences. Furthermore, the determination of other key thresholds and control coefficients within the system is based on the following: by collecting residual fluctuation sequences under normal conditions from historical operating data, three times the maximum amplitude is calculated as the set fluctuation threshold, which is experimentally measured to be 0.05; based on the inflection point where the energy residual no longer decreases when the variational mode decomposition converges, the convergence threshold is set... The sensitivity adjustment coefficient is set to 0.01%, and the reference standard deviation is set to 1.2. These two values ​​were determined through a grid search experiment that balances response speed and noise suppression effect. The safety limit for the rate of temperature change is set to 2.5°C per second to intercept sudden fly-point disturbances of non-physical nature. The dead zone deviation threshold is set to 0.5%, and the deviation trigger threshold is set to 0.2°C. These two values ​​are set based on the machining accuracy and minimum physical dead zone of the electro-hydraulic drive device of the fuel supply valve to ensure that the control loop can maintain high-precision control without causing frequent reciprocating fatigue of the valve.

[0051] To address control misjudgments caused by intermittent signal fly-throughs from the heating furnace sensors, the execution procedure of the sudden change isolation gating unit is refined into data flow consistency judgment logic. Within each control cycle, the system collects measurements from five consecutive sampling points of the outlet temperature. to And calculate the absolute value sequence of differences between adjacent sampling points. If three consecutive differences Exceeding the safety limit for the rate of temperature change And at this time, the transient dynamic prediction error When the absolute value is in the abnormal amplitude range, the gating unit immediately determines that there is a non-thermodynamic abrupt disturbance and locks the corresponding feedback correction bias to 0. Once the gain coefficient returns to the normal range, the gating unit will unlock again. This logic reduces the impact of measurement noise on the prediction iterative closed loop by discretizing the characteristics of the measurement sequence, thereby maintaining the continuous stability of the closed-loop control command.

[0052] Example 5

[0053] Before the shale oil field heaters are put into operation, the system performs a pipeline hysteresis parameter calibration procedure based on offline measured data to establish a physical benchmark for closed-loop control. The calibration procedure requires the heater to operate at 50% rated flow rate in steady state, with the control host sending a step control signal with an amplitude of 10% to the electro-hydraulic actuator of the fuel supply valve. At the same time, a temperature monitoring unit with a sampling frequency of 20Hz records the time delay from the occurrence of the step to the slope reaching 5% at the heater outlet temperature, and defines this time delay as the system's basic physical hysteresis constant. To address the increased execution lag caused by equipment wear or internal coking during furnace operation, the system implements an automatic lag calibration procedure, with the main control unit reading the current valve position feedback value of the fuel supply valve in real time. With controller output command value Calculate the instantaneous position error ,when The deviation is within the preset dead zone threshold for 50 consecutive sampling periods. In addition, the mean of the predicted residuals for the outlet temperature Exceeding the deviation trigger threshold At that time, the calibration program initiates the compensation operator correction, and the updated phase compensation bias is... The calculation formula is: ,in, This is the initial phase compensation bias of the system. This is an adaptive adjustment factor based on error accumulation. The control deviation caused by the decline in valve mechanical response is eliminated by online correction of the command phase, which is obtained through historical abrupt disturbance experiments and calibration.

[0054] The calibration procedure sets the following engineering criteria: when and When the rate of change is less than 0.01 s / cycle, the system determines that the optimal working window has been reached and automatically terminates the compensation operator correction. If the calculated... The system triggers an abnormal state warning and outputs a maintenance work order for the valve drive device to the management terminal to prevent blind execution of compensation logic under degraded equipment performance. The above calibration procedure establishes a direct quantitative correlation between the physical characteristics of the electro-hydraulic actuator and the control algorithm instructions, eliminating the risk of closed-loop control accuracy degradation caused by equipment wear. Under a pipeline pressure of 1.5 MPa, the overshoot deviation of the furnace outlet temperature converged from 2.8℃ to within 0.6℃ through this compensation procedure, proving the effectiveness of the calibration procedure in improving the stability of predictive control.

[0055] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0056] Finally, it should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A closed-loop prediction method for temperature at shale oilfields, characterized in that, Includes the following steps: Step S101: Obtain the current temperature sampling data of the furnace outlet and the fluid volume flow rate data of the furnace feed. Step S102: Call the variational mode decomposition model, and adjust the decomposition scale of the variational mode decomposition model online based on the dynamic residual energy entropy index obtained by asynchronous empirical mode decomposition of the current temperature sampling data, so as to decouple the current temperature sampling data into monotonically variable main trend components and multi-scale time-frequency disturbance components to isolate random noise. Step S103: Input the main trend component and the multi-scale time-frequency perturbation component into the multivariate time-series cascaded prediction network, use the one-dimensional convolutional layer to extract the local time-frequency feature map, and use the long short-term memory network layer to capture the long-term time evolution trend, and output the initial predicted value of the furnace outlet temperature. Step S104: Run the flow adaptive window scale update program, calculate the time-domain phase discrepancy number caused by flow velocity changes based on fluid volume flow rate data and the inherent volume constant of the pipeline network, reconstruct the upper limit of the storage capacity of the circular queue inside the sliding time-domain data matrix to align the physical time phase of the historical comparison sequence in situ when the flow velocity changes, and call the trend consistency identification operator to apply first-order discrete difference operation to the historical comparison sequence to generate the first symbol vector and the second symbol vector respectively. Generate the trend consistency index through point-by-point same sign verification. The mutation isolation gating unit corrects the initial prediction value according to the mapping relationship between the trend consistency index and the transient prediction error to output the target prediction temperature. Use a digital comparator to perform difference comparison operation between the target prediction temperature and the process temperature control target value to output the control valve action increment command, and send it to the electro-hydraulic drive device of the fuel supply valve to adjust the valve opening.

2. The closed-loop prediction method for temperature at shale oilfields according to claim 1, characterized in that, Step S102 includes the following sub-steps: Step S1021, perform empirical mode decomposition on the current temperature sampling data, and calculate the envelope spectrum entropy value of the obtained residual sequence as a dynamic residual energy entropy index; Step S1022, if the difference between the envelope spectrum entropy value and the set entropy benchmark value exceeds the set fluctuation threshold, determine that the current temperature sampling data contains non-stationary random noise and trigger the decomposition scale increment signal; Step S1023, increase the number of mode decompositions of the variational mode decomposition model according to the decomposition scale increment signal until the energy change rate of the residual sequence is lower than the set convergence threshold.

3. The closed-loop prediction method for temperature at shale oilfields according to claim 1, characterized in that, Step S103 includes the following sub-steps: Step S1031, the main trend component and the multi-scale time-frequency perturbation component are matrix-concatenated to construct a multivariate parallel input matrix; Step S1032, the multivariate parallel input matrix is ​​fed into a one-dimensional convolutional layer, and a sliding window convolution kernel of a set size is used to perform feature sliding window scanning along the time axis to extract local time-frequency feature maps; Step S1033, the local time-frequency feature maps are flattened and input into a long short-term memory network layer, and the forget gate, input gate and output gate gate in the layer are used to update and decay the historical state to capture the long-term time evolution trend and output the initial prediction value.

4. The closed-loop prediction method for temperature at shale oilfields according to claim 1, characterized in that, The step S104, which involves running the flow adaptive window scale update program and aligning the physical time phase of the historical comparison sequence in situ, includes the following sub-steps: Step S1041, read the fluid volumetric flow rate data in real time from the data bus, and calculate the corresponding time-domain delay deviator number under the current flow rate condition by calling the geometric transfer function based on the inherent volume constant of the pipeline network; Step S1042, reconstruct the upper limit of the storage capacity of the circular queue inside the sliding time-domain data matrix based on the time-domain delay deviator number, and dynamically increase or decrease the storage length of the circular queue to align the physical time phase of the historical comparison sequence and the current temperature sampling data in situ when the flow rate changes.

5. The closed-loop prediction method for temperature at shale oilfields according to claim 1, characterized in that, The step S104, which calls the trend consistency identification operator to generate a trend consistency index through point-by-point same sign verification, includes the following sub-steps: Step S1043, applying first-order discrete difference operation to the historical comparison sequence to generate a first symbol vector and a second symbol vector respectively; Step S1044, performing point-by-point same sign verification on the first symbol vector and the second symbol vector and performing time-domain point multiplication and accumulation operation within a set time sliding window to generate a trend consistency index.

6. The closed-loop prediction method for temperature at shale oilfields according to claim 1, characterized in that, Step S104, which uses a digital comparator to perform a difference comparison between the target predicted temperature and the process temperature control target value to output an incremental control valve action command, includes the following sub-steps: Step S1045, inputting the target predicted temperature to the first input terminal of the digital comparator, and inputting the process temperature control target value read from the main control memory to the second input terminal of the digital comparator; Step S1046, using the digital comparator to perform a difference comparison between the target predicted temperature and the process temperature control target value, outputting the corresponding incremental control valve action command, and sending it in real time to the electro-hydraulic drive device of the fuel supply valve through the digital output port.

7. The closed-loop prediction method for temperature at shale oilfields according to claim 1, characterized in that, Step S101 includes the following sub-steps: Step S1011, collect real-time temperature data of the heating furnace outlet to generate current temperature sampling data; Step S1012, synchronously read the real-time volumetric flow rate of the heating furnace feed through the flow meter data bus to generate fluid volumetric flow rate data.

8. The closed-loop prediction method for temperature at shale oilfields according to claim 1, characterized in that, After step S104, the following steps are also included: Step S105, in each closed-loop control cycle, the target predicted temperature is compared with the set safety upper limit temperature by amplitude. If the target predicted temperature is greater than the safety upper limit temperature, an emergency fuel cut-off command is triggered and the control valve action increment command is forcibly rewritten as a safety valve closing command to close the fuel supply valve.

9. The closed-loop prediction method for temperature at shale oilfields according to claim 1, characterized in that, The entire prediction and control process runs spontaneously in a closed loop with a fixed control cycle. At the beginning of each control cycle, the latest current temperature sampling data is reacquired and the corresponding control valve action increment command is recalculated. After the material flow fluctuation triggers adaptive feedback adjustment, the furnace outlet temperature is stabilized within the error control range set by the process.

Images