Dynamic liquid level prediction method and device, electronic device and storage medium
Through neural network model and data processing technology, the problem of complexity in the relationship between the dynamic fluid level and the fluid production volume of the oil well is solved, dynamic fluid level prediction and steady state judgment are realized, and oil production efficiency and equipment life are improved.
Patent Information
- Application Number
- CN202210092824.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-26
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-01-26
AI Technical Summary
The existing technology cannot accurately calculate the complex change relationship between the hydraulic level and the liquid production of the oil well, resulting in poor operation of the oil well and affecting the oil production efficiency.
The neural network model is used to combine historical data, and predict the dynamic fluid level through data alignment, state decomposition and model update, determine whether the new data is steady-state, and retrain the model to improve prediction accuracy.
The accuracy and steady-state judgment of dynamic fluid level prediction are achieved, the optimization capability of oil well operation is improved, the life of the oil production system is extended, and energy consumption is reduced.
Smart Images

Figure CN114444392B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an artificial intelligence analysis algorithm framework for oil well operating conditions during oil production in the petroleum industry. Background Art
[0002] During the operation of oil wells, affected by various factors, the operation will deviate from the optimal state, resulting in eccentric wear, shortened pump inspection cycle, decreased production, increased energy consumption and other problems. These problems cause great economic losses to oil field production, and it is urgent to optimize and transform the oil production system. At present, the basis for optimizing the benefits of the oil production system is the optimized operation of a single well. The control principle of the optimized operation of a single well is to achieve the effect of supply and drainage coordination, efficiency improvement and life extension by monitoring the dynamic liquid level and liquid production, and adjusting the number of strokes and strokes in real time. The core of the optimized operation of a single well is to obtain accurate dynamic liquid level and liquid production in real time. Therefore, in the oil production of oil wells in oil fields, the dynamic liquid level depth is a very important parameter.
[0003] At present, the oil industry mainly adopts the method of engineering principle modeling to calculate the dynamic liquid level and liquid production. The specific process is to collect real-time oil well operating data such as displacement, load, back pressure, current, voltage, temperature, etc. through the digital oilfield information system. Then, according to the changing relationship between the dynamic liquid level, liquid production and this set of operating parameters, an equation is fitted for calculation.
[0004] The current situation in the oilfield industry is that engineering modeling cannot accurately calculate the complex relationships of the above changes, resulting in inaccurate results and affecting the optimal operation of oil wells. Summary of the invention
[0005] The present invention is made in view of the above problems in the prior art, and is used to solve one or more problems existing in the prior art.
[0006] According to one aspect of the present invention, a method for predicting a dynamic liquid level is provided, characterized in that it comprises the following steps: obtaining a plurality of dynamic liquid level parameter states and a model set {m i , F i (x)}, where m i represents the i-th state of the dynamic liquid surface parameter, F i (x) represents the neural network model related to the i-th state, and x is the dynamic liquid surface parameter data related to the i-th state; calculate the new data X t With each state m i The similarity of each data in the data set under the same condition, if the maximum similarity is greater than the preset minimum similarity, the new data X t Add the data in the corresponding state and retrain the model F i (x), obtain multiple updated states and model sets {m i , F i(x)}, retrain the model F i (x), perform convergence calculation and judge the new data X t whether the corresponding state after being added is a steady state; when the corresponding state is a non-steady state, dividing the new data into a new state; and using the retrained model to predict the dynamic liquid level.
[0007] According to another aspect of the present invention, a dynamic liquid level prediction device is provided, comprising: a state model acquisition unit, which obtains a plurality of dynamic liquid level parameter states and a model set {m i , F i (x)}, where m i represents the i-th state of the dynamic liquid surface parameter, F i (x) represents the neural network model related to the i-th state, and x is the dynamic liquid surface parameter data related to the i-th state; the update unit calculates the new data X t With each state m i The similarity of each data in the data set under the same condition, if the maximum similarity is greater than the preset minimum similarity, the new data X t Add the data in the corresponding state and retrain the model F i (x), obtain multiple updated states and model sets {m i , F i (x)}, the update unit retrains the model F i (x), perform convergence calculation and judge the new data X t Whether the corresponding state after being added is a steady state, when the corresponding state is a non-steady state, the new data is divided into a new state; the prediction unit uses the retrained model to predict the dynamic liquid level.
[0008] According to yet another aspect of the present invention, an electronic device is provided, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the method described in the present invention.
[0009] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a device control program is stored. When the device control program is executed by a processor, the method described in the present invention is implemented.
[0010] According to the embodiment of the present invention, the dynamic liquid surface can be predicted well. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] The present invention can be better understood with reference to the accompanying drawings, which are merely illustrative and are not intended to limit the scope of protection of the present invention.
[0012] Figure 1 A schematic flow chart of a method for predicting the dynamic liquid level depth according to an embodiment of the present invention is shown.
[0013] Figure 2 A schematic flow chart of a method for performing data alignment according to an embodiment of the present invention is shown.
[0014] Figure 3 A schematic diagram of state decomposition according to an embodiment of the present invention is shown.
[0015] Figure 4 A flow chart of a method for establishing a model for dynamic liquid level prediction according to an embodiment of the present invention is shown.
[0016] Figure 5 and Figure 6 A schematic diagram of a neural network architecture configuration according to an embodiment of the present invention is shown.
[0017] Figure 7 A schematic block diagram of a dynamic liquid level depth prediction device according to an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0018] The specific embodiments of the present invention are described below in conjunction with the accompanying drawings. These descriptions are exemplary and are intended to enable those skilled in the art to implement the embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. The description does not describe any content that is indispensable for actual implementation but irrelevant to understanding the present invention.
[0019] Figure 1 A schematic flow chart of a method for predicting the dynamic liquid level depth according to an embodiment of the present invention is shown.
[0020] like Figure 1 As shown, according to one embodiment of the present invention, firstly, in step S100, dynamic liquid level parameters and dynamic liquid level historical data related to the dynamic liquid level are obtained in chronological order. This can be obtained, for example, by collecting existing historical data. The previously accumulated database can be used. The dynamic liquid level parameters are roughly divided into operating electrical parameters (current, voltage, displacement, load, etc.) of the oil well pumping unit and sensor parameters (back pressure, temperature, etc.) at the wellhead of the oil well. In this article, these parameters are collectively referred to as dynamic liquid level parameters. Collect these parameters and the corresponding dynamic liquid level depth in chronological order.
[0021] Then, in step S200, the data are aligned according to the time when they were generated. Because the production condition data are collected by different devices, it is necessary to consider the production and collection process, and combine the data according to time to prepare for subsequent modeling.
[0022] Figure 2FIG. 1 is a schematic flow chart of a method for performing data alignment according to an embodiment of the present invention. Figure 2 As shown, according to one embodiment of the present invention, after the data from the pumping units of each oil well are summarized, they are merged according to the acquisition time order of each dynamic liquid level parameter. For example, the acquisition time order of the current is in the order of seconds, that is, the current data can be acquired at intervals of seconds. If the time order is consistent, then the data is merged. According to one embodiment, intelligent filling is performed to fill in the missing working condition data and dynamic liquid level data. After filling, the data is merged.
[0023] According to one implementation, the historical data related to the dynamic liquid level are arranged according to the dynamic liquid level parameters of the minimum acquisition time, and the parameters of the larger time acquisition level are interpolated and filled according to the production cycle of the corresponding oil pump. For example, there is one line of load every 30 minutes, one line of dynamic liquid level every 2 hours, and one line of current every 1 minute, and the dimension of current is taken as the shortest acquisition interval). According to the production cycle of the oil pump (that is, the process from the lowest point to the highest point and back to the lowest point), the data of the dynamic liquid level and the data of the load are interpolated to obtain the data corresponding to the sampling interval of the current, forming a time-consistent data including the current, dynamic liquid level, and load.
[0024] According to this method, the existing data can be better utilized and interpolated according to the production cycle of the oil pumping unit, thereby preventing the total amount of data from being too large and ensuring the accuracy of the prediction.
[0025] After data merging, each liquid level parameter and dynamic liquid level data can be merged into one row according to the corresponding time, so that a data table after data alignment can be obtained.
[0026] Then, in step S300, data state decomposition is performed to eliminate data ambiguity. According to one embodiment, a projection transformation method can be used to perform state decomposition.
[0027] Figure 3 FIG. 1 is a schematic diagram of state decomposition according to an embodiment of the present invention. Taking a displacement of a dynamic liquid level parameter as an example, for example, assuming that the historical data of the displacement a1 is a1, a2, ...a n , define the projection transformation function as H(x)
[0028] (1) Assume that there are m states (e.g., the liquid level rises, the liquid level drops, and the liquid level fluctuates): replace a1, a2, ...a n Each data is randomly assigned to a state (1, 2...m). The data of a state x is {a x}. Here {a x} can be a collection of data, such as a1, a2, ...a k, k is a positive integer less than n.
[0029] (2) Calculate the indicator for each state: y x =H({a x})
[0030] (3) Calculate the rationality index value of the decomposition: For example, let y1, y x2 ...y x The variance of is the largest, and the index value (variance) of this decomposition is calculated.
[0031] (4) Use the optimization algorithm to continuously adjust m and repeat steps (1) to (3) until the indicator cannot be improved, thereby obtaining the corresponding {x i , m i}.
[0032] i represents the number of all historically collected data, which are arranged in order from the earliest to the latest time. i Indicates the dynamic liquid level parameter at the i-th time point. For example: there are 100 million historical data, the first data is called x1, m i Indicates this row of data x i Which state it belongs to. For example: a has three states (dynamic liquid level rising, dynamic liquid level falling, dynamic liquid level fluctuating), m1 represents the state of dynamic liquid level rising.
[0033] Then, in step S400, a dynamic liquid level prediction model is established. Any deep neural network based on gradient descent can be used to establish a dynamic liquid level prediction model for preliminary prediction.
[0034] Figure 4 A flow chart of a method for establishing a model for dynamic liquid level prediction according to an embodiment of the present invention is shown.
[0035] like Figure 4 As shown, according to one embodiment of the present invention, a method for establishing a dynamic liquid level prediction model includes the following steps:
[0036] (1) Steps for determining the training set and test set: Obtain dynamic liquid level parameter data (production condition data) x i , run monitoring data y i (e.g. dynamic liquid level), state number m i , according to the time splicing into {x i , m i ,y i}, according to different states m i , group the data. For example, there are three states, and the data group is m0: {x i ,y i}、m1:{x i ,y i}、m2:{x i ,y i Then, the grouped data is divided into a training set and a test set according to a certain ratio (e.g. 8:2);
[0037] (2) Preliminary model setting steps: The input of the model is {x i , m i}, the output is {y i}. The loss function of the model is set to Obj(θ)=L(θ)+Ω(θ)+Z(θ), where L(θ) represents the error function. According to one embodiment, the error function L(θ) is the mean square error (MSE)
[0038]
[0039] Where n represents the total number of samples, yi represents an actual dynamic liquid surface, Represents the dynamic liquid surface predicted by the model. MSE reflects the accuracy of the model prediction. The more accurate the prediction, the smaller the MSE. When the model is trained, the model parameters will be adjusted according to the direction with the smallest MSE.
[0040] Ω(θ): represents the penalty for complex models. A loss equation can generally only solve one problem. The L(θ) designed above is used to ensure the model's fit, while we also need to ensure the generalization of the model. If the model's complexity is not controlled, the model will have a good prediction effect on historical data, but the accuracy of actual predictions in the future will drop significantly.
[0041] Z(θ): Penalizes models that do not conform to expert experience. The design of L(θ) and Ω(θ) is based on traditional data-driven scenarios. In the oil extraction scenario, except for data missing and data anomalies, there are obvious formulaic equality or inequality relationships in the data.
[0042] For example
[0043] Dynamic liquid level Load Current Power average Temperature, pressure
[0044] It means influence. The importance of the load, current, power average, temperature, and pressure in the generated model should also conform to this rule. In addition, in oil production, as the number of strokes increases, the dynamic liquid level will normally have a downward trend in the long run, and the prediction results of the model must also conform to this rule.
[0045] The loss function Z(θ) here is generally designed by combining expert experience and universal laws reflected in historical data.
[0046] (3) Neural network structure optimization steps: Use algorithms such as the neural architecture automatic search algorithm to automatically optimize the neural network architecture (network width, number of hidden layers, number of neurons, activation function, etc.)
[0047] Figure 5 and Figure 6 A schematic diagram of a neural network architecture arrangement according to an embodiment of the present invention is shown. Figure 5 In the case of the data relationship shown, for example, Figure 6 The network framework is constructed and the neural network architecture is optimized using the neural architecture automatic search algorithm. Figure 6 In the equation, x1: temperature; x2: pressure; x3: average power; x4: current; x5: load. The model can be obtained by learning from data or designed and input in advance.
[0048] (4) Parameter optimization step: Use optimization algorithms such as HPO, Bayesian, and random evolution to optimize the remaining hyperparameters (learning rate, number of iterations, batch size, etc.) of the preliminary model in step (2).
[0049] (5) According to the optimized parameters found, in the total amount of {x i , m i ,y i}, train and obtain the final neural network model F(x).
[0050] During actual operation, due to changes in actual production conditions, the currently obtained conditions may belong to a certain historical condition or a new condition may have emerged. The parameters of different condition models need to be recalibrated, otherwise the prediction accuracy will decrease.
[0051] Therefore, in step S500, the neural network model F(x) is updated using the new data.
[0052] According to one embodiment, for the newly acquired data in actual operation, the default state is firstly the closest to the previous historical data. The new data X is calculated using a clustering algorithm. t With each state m in history i The data under (with state m i The clustering algorithm is used to determine a state belonging to the history, and the model parameters in this state are updated and retained.
[0053] For example, the state and model set obtained from the historical data is {m i , F i (x)}, set the minimum similarity, and use the clustering algorithm to calculate the new data X t With each state m in historyi If the maximum similarity is greater than the minimum similarity set, the new data X t Add the data in the corresponding state and retrain the model. If the maximum similarity is less than the set minimum similarity, the new data is divided into a new state. i (x) refers to the model in the i-th state.
[0054] Assume the corresponding state is m1, the historical model F1(x) = F1 + X1
[0055] The new model F1(x)=F1+(X1+X t ).
[0056] With the self-optimization update after the actual operation of the model, many states of the oil well and the corresponding model sets {m i , F i (x)},
[0057] Among them, F i (x) = F i +X i , F i refers to the architecture of the model of the i-th state.
[0058] After updating the model, a convergence calculation is performed to determine whether the state is a steady state. If it is a steady state, a unique corresponding functional relationship model for the state is established.
[0059] According to one embodiment, whether or not a steady state exists may be determined as follows.
[0060] The data of a production cycle {x1, x2, ...x t} is definitely a state, which may be one of the historical states or a completely new state. t When the rationality index value of the decomposition is calculated, the convergence is determined according to the optimal direction. The state decomposition method in step S300 can be referred to. If it cannot be divided into more states, it is converged. Convergence means it belongs to this state, and non-convergence means it does not belong to this state.
[0061] If we judge the latest data {x1, x2, ...x t} no longer belongs to this state t If the state is not stable, a new state is established for the new data.
[0062] According to an embodiment of the present invention, the similarity between new data and historical data classified into a certain state is first determined, and then whether each state is steady after updating is determined, so that the model can be better updated and achieve better prediction effect.
[0063] According to an embodiment of the present invention, the method further includes a step of determining a sub-state.
[0064] For each state m i , the overall best prediction model architecture F i , the actual parameters of the model are trained using all the data in this state. For some of the data, a more accurate model can be obtained. For example, suppose that for a certain state m1, the predicted dynamic liquid level model found is y1=0.15x1+0.15x2 (Formula 1). Depending on the model parameters, the model can be y1=0.1x1+0.2x2 (Formula 2), y1=0.2x1+0.3x2 (Formula 3)... etc. The prediction accuracy of these models is different. If the variant model (Formula 2) has a higher accuracy on some data than the overall model (Formula 1), these data are a sub-state under the state m1.
[0065] Optimization algorithms such as evolutionary strategies and simulated annealing can be used to determine the minimum amount of data for a sub-state and the range of variation of model parameters, so that the distribution of sub-states and the model parameters under the sub-states (the only corresponding functional relationship model) can be automatically obtained.
[0066] According to one embodiment, a clustering algorithm different from the state decomposition method is used to find multiple sub-states and corresponding data sets {m ij , x ij}, and then train based on the data set to obtain a sub-state model with higher prediction accuracy than the model of this state.
[0067] According to one embodiment, the parameters of the overall model architecture of each state are adjusted to predict part of the data in the data set corresponding to the state. A model with a higher accuracy than the overall model architecture is obtained as a sub-state model, and the part of the data is used as the data set corresponding to the sub-state.
[0068] In the step of determining the sub-states, multiple sub-states and corresponding data sets {m ij , x ij}, where m ij represents the jth substate of the i-th state of the dynamic liquid surface parameter, x ij Represents the data set related to the sub-state, and uses the data set under the sub-state to build the corresponding model f ij (x).
[0069] Calculate new data X t With each state m i When the similarity of the data under t With each sub-state mij If the maximum similarity is greater than the minimum similarity set, the new data X t Add the data under the corresponding sub-state.
[0070] When the model is actually used, the recently acquired new data is used to determine which sub-state the current production is in, and then the model parameters in the sub-state are used for prediction.
[0071] Figure 7 A schematic block diagram of a dynamic liquid level depth prediction device according to an embodiment of the present invention is shown.
[0072] like Figure 7 As shown, according to an embodiment of the present invention, the dynamic liquid surface depth prediction device includes: a state model acquisition unit 700, which obtains multiple dynamic liquid surface parameter states and model sets {m i , F i (x)), where m i represents the i-th state of the dynamic liquid surface parameter, F i (x) represents the neural network model associated with the i-th state, and x is the dynamic liquid surface parameter data associated with the i-th state; the updating unit 800 calculates the new data X t With each state m i The similarity of each data in the data set under the same condition, if the maximum similarity is greater than the preset minimum similarity, the new data X t Add the data in the corresponding state and retrain the model F i (x), obtain multiple updated states and model sets {m i , F i (x)}. The updating unit 800 retrains the model F i (x), perform convergence calculation and judge the new data X t Whether the corresponding state after being added is a steady state, when the corresponding state is a non-steady state, the new data is divided into a new state; the prediction unit 900 uses the retrained model to predict the dynamic liquid level.
[0073] According to one embodiment, the dynamic liquid level depth prediction device further includes a sub-state acquisition unit, which finds multiple sub-states and corresponding data sets {m ij ,x ij}, where m ij represents the jth substate of the i-th state of the dynamic liquid surface parameter, x ij Represents the data set related to the sub-state, and uses the data set under the sub-state to build the corresponding model f ij(x), a sub-state refers to a state that can obtain a higher prediction accuracy than the state in which it is located based on part of the data in the data set of the state in which it is located. In this case, the update unit 800 calculates the new data X t With each state m i When the similarity of the data under t With each sub-state m ij If the maximum similarity is greater than the minimum similarity set, the new data X t Add the data under the corresponding sub-state.
[0074] According to one embodiment, the updating unit 800 determines whether the i-th state is a steady state as follows:
[0075] (1) Assume that a dynamic liquid level parameter has m states: the historical data a1, a2, ...a n Each data is randomly assigned to a state (1, 2...m), and the data of a state x is {a x};
[0076] (2) Calculate the indicator for each state: y x =H({a x}), H(x) is the projection transformation function;
[0077] (3) Calculate the rationality index value of the decomposition;
[0078] (4) Use the optimization algorithm to continuously adjust m and repeat steps (1) to (3) until the indicator cannot be improved, thus obtaining {x i , m i};
[0079] Among them, i represents the number of all historically collected data, which are arranged in order from the earliest to the latest time, and x i Represents the dynamic liquid level parameter data at a certain time point, m i Indicates this row of data x i When the state it belongs to cannot be divided into more states, it is judged to be convergent, that is, a steady state.
[0080] According to one embodiment, the state model acquisition unit 700 uses a state decomposition method to obtain multiple dynamic liquid level parameter states and model sets {m i , F i (x)}, the sub-state acquisition unit uses a clustering algorithm different from the state decomposition method to find multiple sub-states and corresponding data sets {m ij ,x ij}.
[0081] According to one embodiment, when the prediction unit 900 uses the retrained model to predict the dynamic liquid level, it first uses the most recently acquired new data to determine which sub-state the current production is in, and then uses the model parameters in the sub-state to perform prediction.
[0082] According to one embodiment, the dynamic liquid level depth prediction device also includes a time alignment unit, which arranges historical data related to the dynamic liquid level according to the dynamic liquid level parameters of the minimum acquisition time acquisition level, and interpolates and fills the parameters of the larger time acquisition level according to the production cycle of the corresponding oil pump.
[0083] According to one embodiment, the sub-state acquisition unit finds multiple sub-states and corresponding data sets {m ij ,x ij}, and then train based on the data set to obtain a model with higher prediction accuracy than the model of this state as the sub-state model, or adjust the parameters of the model of each state, and predict part of the data in the data set corresponding to the state, and obtain a model with higher accuracy than the overall model architecture as the sub-state model, and use the part of the data as the data set corresponding to the sub-state.
[0084] Due to the internal complexity of currently commonly used models such as neural networks, it is usually difficult to know why such a prediction value is obtained after making a prediction result. Oil well production management often requires controlling the motor based on the predicted value. If there is no controllable explainability, it is difficult to apply it in the actual production process.
[0085] The present invention can provide prediction results in the model usage analysis stage and, based on the specified control objectives, give adjustment suggestions, and can implement sensitivity analysis (Sensitivity Analysis) and gradient-based methods (Gradient-based Methods) to transform the black box model into a white box model.
[0086] For example, in the production process of an oil well, according to the production conditions of the underground oil layer and the changes in the formation oil pressure, it is hoped that the daily liquid production will be maximized while keeping the dynamic liquid level constant. The predicted dynamic liquid level is obtained based on the existing stroke and stroke frequency. The dynamic liquid level is predicted again by slightly increasing the stroke and stroke frequency until the predicted dynamic liquid level drops to an acceptable limit. At this time, the stroke and stroke frequency are the maximum acquisition frequency that can be achieved at present, that is, the maximum daily liquid production.
[0087] According to the embodiments of the present invention, methods such as data clustering based on data feature analysis, sub-state decomposition based on model analysis, and model adaptive self-updating are used to solve the problem of low model prediction accuracy caused by frequent changes in operating conditions.
[0088] According to the implementation mode of the present invention, based on the deep reasoning of the model, real-time prediction is achieved, and the dynamic liquid level is adjusted in time to restore the dynamic liquid level to a normal range before the abnormal dynamic liquid level causes problems.
[0089] It is easy for a person skilled in the art to understand that the apparatus of the present invention can be understood by using the above description of the method.
[0090] Those skilled in the art will readily understand that the method of the present invention may also include other steps corresponding to the functions performed by the device of the present invention. The above steps may also be simplified.
[0091] The labels for units and steps in the present invention are only for convenience of explanation and do not represent the order of execution unless there is contrary description in the context.
[0092] Those skilled in the art should understand that the above-mentioned units can be implemented by software or special hardware, such as a field programmable gate array, a single-chip microcomputer, or a microchip, or can also be implemented by combining software with hardware.
[0093] The present invention also provides an electronic device, comprising: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the method of the present invention.
[0094] The present invention also relates to a computer software, which can implement the method of the present invention when executed by a computing device (such as a single chip microcomputer, a computer, a CPU, etc.).
[0095] The present invention also relates to a computer software storage device, such as a hard disk, a floppy disk, a flash memory, etc., wherein the computer software storage device stores the above-mentioned computer software.
[0096] The description of the method or step of the present invention can be used to understand the description of the unit or device, and the description of the unit or device can also be used to understand the method or step of the present invention.
[0097] The above description is only illustrative and is not intended to limit the scope of protection of the present invention. Any changes and substitutions within the scope of the claims of the present invention are within the scope of protection of the present invention.
Claims
1. A method for predicting a dynamic liquid level, characterized in that: The steps include: From the historical data related to the dynamic liquid surface, multiple dynamic liquid surface parameter states and model sets {m i ,F i (x)}, where m i represents the i-th state of the dynamic liquid surface parameter, F i (x) represents the neural network model associated with the i-th state, and x is the dynamic liquid surface parameter data associated with the i-th state; Calculate new data X t With each state m i The similarity of each data in the data set under the following formula, if the maximum similarity is greater than the preset minimum similarity, the new data X t Add the data in the corresponding state and retrain the model F i (x), obtain multiple updated states and model sets {m i ,F i (x)}, retrain the model F i (x), perform convergence calculation and judge the new data X t whether the corresponding state after adding is a steady state, and when the corresponding state is a non-steady state, dividing the new data into a new state; as well as Using the retrained model for dynamic liquid level prediction, Wherein, the method further comprises: Find multiple sub-states and corresponding data sets for each state {m ij , x ij }, where m ij represents the jth substate of the i-th state of the dynamic liquid surface parameter, x ij Represents the data set related to the sub-state, and uses the data set under the sub-state to build the corresponding model f ij (x), a sub-state refers to a state that can obtain a higher prediction accuracy than the state it is in based on part of the data in the data set of the state it is in. Among them, calculate the new data X t With each state m i When the similarity of the data under t With each sub-state m ij If the maximum similarity is greater than the minimum similarity set, the new data X t Add the data under the corresponding sub-state.
2. The method according to claim 1, characterized in that Whether the i-th state is a steady state is determined as follows: (1) Assume that a dynamic liquid level parameter has m states: the historical data a1, a2, ...a n Each data is randomly assigned to a state (1,2...m), and the data of a state x is {a x }; (2) Calculate the indicator for each state: y x =H({a x }), H(x) is the projection transformation function; (3) Calculate the rationality index value of the decomposition; (4) Use the optimization algorithm to continuously adjust m and repeat steps (1) to (3) until the indicator cannot be improved, thus obtaining {x i ,m i }; Among them, i represents the number of all historically collected data, which are arranged in order from the earliest to the latest time, and x i Represents the dynamic liquid level parameter data at a certain time point, m i Indicates this row of data x i When the state it belongs to cannot be divided into more states, it is judged to be convergent, that is, a steady state.
3. The method according to claim 1, characterized in that: The state decomposition method is used to obtain multiple dynamic liquid surface parameter states and model sets {m i ,F i (x)}, using a clustering algorithm different from the state decomposition method to find multiple sub-states and corresponding data sets {m ij , x ij }.
4. The method according to claim 3, characterized in that When using the retrained model to predict the dynamic liquid level, the most recently acquired new data is first used to determine which sub-state the current production is in, and then the model parameters in the sub-state are used to make predictions.
5. The method according to claim 1, characterized in that The method also includes a time alignment step, arranging historical data related to the dynamic liquid level according to each dynamic liquid level parameter of the minimum acquisition time acquisition level, and interpolating and filling parameters of the larger time acquisition level according to the production cycle of the corresponding pumping unit.
6. The method according to claim 1, characterized in that Find multiple sub-states and corresponding data sets for each state {m ij , x ij }, and then train based on the data set to obtain a model with higher prediction accuracy than the model of this state as the sub-state model.
7. A dynamic liquid level prediction device, characterized in that: include: The state model acquisition unit obtains multiple dynamic liquid surface parameter states and model sets {m i ,F i (x)}, where m i represents the i-th state of the dynamic liquid surface parameter, F i (x) represents the neural network model associated with the i-th state, x is the dynamic liquid surface parameter data associated with the i-th state, Update unit, calculate new data X t With each state m i The similarity of each data in the data set under the following formula, if the maximum similarity is greater than the preset minimum similarity, the new data X t Add the data in the corresponding state and retrain the model F i (x), obtain multiple updated states and model sets {m i ,F i (x)}, the update unit retrains the model F i (x), perform convergence calculation and judge the new data X t whether the corresponding state after adding is a steady state, and when the corresponding state is a non-steady state, dividing the new data into a new state; The prediction unit uses the retrained model to predict the dynamic liquid level. The dynamic liquid level prediction device further includes a sub-state acquisition unit, which finds multiple sub-states and corresponding data sets {m ij , x ij }, where m ij represents the jth substate of the i-th state of the dynamic liquid surface parameter, x ij Represents the data set related to the sub-state, and uses the data set under the sub-state to build the corresponding model f ij (x), a sub-state refers to a state that can obtain a higher prediction accuracy than the state it is in based on part of the data in the data set of the state it is in. The updating unit calculates the new data X t With each state m i When the similarity of the data under t With each sub-state m ij If the maximum similarity is greater than the minimum similarity set, the new data X t Add the data under the corresponding sub-state.
8. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the method according to any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a device control program, and when the device control program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.
