Parameter optimization method and related apparatus

Abstract

Provided in the present application are a parameter optimization method and a related apparatus, which are applied to the field of computers, and are used for optimizing control parameters of an objective optimization function on the basis of a prediction result obtained from a temporal prediction model performing prediction on time series data, so as to make a result obtained after optimization more suitable for an actual working condition objective. The method comprises: acquiring time series data of a key indicator, wherein the key indicator is used for optimizing control parameters; on the basis of the time series data, using a temporal prediction model to perform prediction, so as to obtain predicted data of the key indicator; determining an objective optimization function on the basis of the temporal prediction model and a target value of the key indicator, wherein the objective optimization function represents the difference between the predicted data of the key indicator and the target value thereof; and optimizing the control parameters in the objective optimization function on the basis of the predicted data of the key indicator, so as to obtain optimized control parameters.

Description

Parameter optimization method and related device

[0001] The present application claims priority from the Chinese patent application No. 202410417654.8 filed on April 8, 2024, and entitled "A parameter optimization method and related device", the whole content of which is incorporated herein by reference. TECHNICAL FIELD

[0002] The present application relates to the field of computers, and in particular to a parameter optimization method and related device. BACKGROUND

[0003] Process industry is the pillar industry and basic raw material industry of China's national economic development, mainly including high energy consumption industries such as petroleum, chemical industry, mining, iron and steel, non-ferrous metal, electric power and building materials. In the process industry, production efficiency can be improved, cost can be reduced, and product quality can be improved by optimizing process parameters.

[0004] Currently, the optimization of process parameters in the process industry is mainly based on proportion integration differentiation (PID) parameter recommendation and standard evolutionary algorithms such as inheritance algorithm and particle swarm algorithm for parameter optimization.

[0005] However, due to the existence of prior knowledge and constraint conditions in the industrial scene, i.e., production requirements, only using standard evolutionary algorithms for process parameter optimization may not achieve the process condition target in the process industry. SUMMARY

[0006] The present application provides a parameter optimization method and related device, which is used for optimizing the control parameters of a target optimization function according to the prediction result of the time series prediction model for predicting time series data, so that the optimized result is more suitable for the actual process condition target.

[0007] Therefore, in a first aspect, the present application provides a parameter optimization method. In the method, firstly, the change of an acquired key indicator can be recorded to obtain time series data of the key indicator. Then, the time series data can be used to predict the key indicator by using a time series prediction model to obtain prediction data of the key indicator, wherein the prediction data can include the values of the key indicator at multiple different future time points, so that the gap between the predicted value of the key indicator and the target value of the key indicator can be determined according to the prediction data. Before determining a target optimization function, the target value of the key indicator can also be obtained. The target optimization function can be determined by combining the target value of the key indicator and the time series prediction model, and the target optimization function can represent the difference between the prediction data of the key indicator and the target value. Furthermore, the control parameters in the target optimization function can be optimized according to the prediction data of the key indicator, so that the target optimization function reaches an optimal value, i.e., the predicted value of the key indicator is as close as possible to the target value, thereby obtaining the optimized control parameters.

[0008] In the embodiments of the present application, the prediction data of the key indicator can be obtained by using the time series prediction model according to the time series data of the key indicator, and the control parameters in the target optimization function can be optimized according to the prediction data of the key indicator. The target optimization function can represent the difference between the prediction data of the key indicator and the target value of the key indicator. Therefore, by adjusting the control parameters in the target optimization function, the value of the target optimization function is adjusted, so that the value of the key indicator under the optimized control parameters is as close as possible to the target value, thereby better adapting to the actual working condition target.

[0009] In a possible implementation, the aforementioned prediction of the prediction data of the key indicator by using the time series prediction model according to the time series data can include: performing frequency domain multi-scale learning on the time series data to obtain a feature representation of the time series data, the feature representation including multi-scale features of the time series data; extracting time series information of the time series data by using time series convolution to obtain time series features of the time series data; and predicting the prediction data of the key indicator by using the time series prediction model according to the feature representation and the time series features.

[0010] In the embodiments of the present application, the features of the time series data can be extracted from both the frequency domain and the time domain, so that the extracted features of the time series data are more abundant, and the prediction result is more accurate.

[0011] In a possible implementation, the aforementioned prediction of the prediction data of the key indicator by using the time series prediction model according to the feature representation and the time series features can include: predicting the prediction data of the key indicator by using a multi-layer perception according to the feature representation and the time series features.

[0012] In a possible implementation, the aforementioned frequency domain multi-scale learning of the time series data to obtain the feature representation of the time series data can include: performing fast Fourier transform on the time series data to obtain frequency domain features of the time series data; performing multi-scale pooling on the frequency domain features to obtain pooled frequency domain features; performing multi-scale feature extraction on the pooled frequency domain features to obtain multi-scale features; and performing inverse fast Fourier transform on the multi-scale features to obtain the feature representation.

[0013] In the embodiments of the present application, the time series data can be subjected to frequency domain multi-scale learning, and multi-scale pooling is performed on the time series data to filter noise in the time series data, thereby enhancing the robustness of the time series prediction model.

[0014] In a possible implementation, the aforementioned determination of the target optimization function according to the time series prediction model and the target value of the key indicator can include: determining a first loss function in the target optimization function according to the time series prediction model and the target value of the key indicator, the first loss function representing a difference between the current value and the target value of the key indicator; and obtaining the target optimization function according to the first loss function and a weight of the first loss function.

[0015] In the embodiments of the present application, the target value of the key indicator is introduced, and the target optimization function can be determined according to the time series prediction model and the target value, so that the optimization of the control parameter is more suitable for the actual working condition target, and the control parameter can be adjusted and optimized according to the change trend of the key indicator.

[0016] In a possible implementation, the aforementioned determination of the first loss function in the target optimization function according to the time series prediction model and the target value of the key indicator can include: determining whether the key indicator meets a preset requirement according to the time series prediction model and the target value; if the preset requirement is not met, determining the first loss function according to a first coefficient and a second loss function, the second loss function being a difference between the current value and the target value of the key indicator; and if the preset requirement is met, determining the first loss function according to a second coefficient and the second loss function, the second coefficient being an inverse of the first coefficient.

[0017] In a possible implementation, the aforementioned optimization of the control parameter in the target optimization function according to the predicted data of the key indicator to obtain the optimized control parameter can include: optimizing the control parameter in the target optimization function according to the predicted data of the key indicator by using an optimization algorithm to obtain the optimized control parameter.

[0018] In a second aspect, the present application provides a parameter optimization device, comprising:

[0019] The acquisition module is configured to acquire time series data of a key indicator, the key indicator being used for optimizing a control parameter.

[0020] The prediction module is configured to perform prediction on the time series data by using a time series prediction model to obtain prediction data of the key indicator.

[0021] The determination module is configured to determine a target optimization function according to the time series prediction model and the target value of the key indicator, the target optimization function representing a difference between the prediction data of the key indicator and the target value.

[0022] The optimization module is configured to optimize the control parameter in the target optimization function according to the prediction data of the key indicator to obtain an optimized control parameter.

[0023] In a possible implementation, the prediction module is specifically configured to: perform frequency domain multi-scale learning on the time series data to obtain a feature representation of the time series data, the feature representation including multi-scale features of the time series data; extract time series information of the time series data by using a time series convolution to obtain time series features of the time series data; and perform prediction by using a time series prediction model according to the feature representation and the time series features to obtain the prediction data of the key indicator.

[0024] In a possible implementation, the prediction module is specifically configured to: perform prediction by using a multi-layer perception according to the feature representation and the time series features to obtain the prediction data of the key indicator.

[0025] In a possible implementation, the prediction module is specifically configured to: perform fast Fourier transform on the time series data to obtain frequency domain features of the time series data; perform multi-scale pooling on the frequency domain features to obtain pooled frequency domain features; perform multi-scale feature extraction on the pooled frequency domain features to obtain multi-scale features; and perform inverse fast Fourier transform on the multi-scale features to obtain the feature representation.

[0026] In a possible implementation, the determination module is specifically configured to: determine a first loss function in the target optimization function according to the time series prediction model and the target value of the key indicator, the first loss function representing a difference between a current value of the key indicator and the target value; and obtain the target optimization function according to the first loss function and a weight of the first loss function.

[0027] In a possible implementation, the determination module is specifically configured to: determine whether the key indicator meets a preset requirement according to the time series prediction model and the target value; if the key indicator does not meet the preset requirement, determine the first loss function according to a first coefficient and a second loss function, the second loss function being a difference between the current value of the key indicator and the target value; and if the key indicator meets the preset requirement, determine the first loss function according to a second coefficient and the second loss function, the second coefficient being an inverse of the first coefficient.

[0028] In a possible implementation, the optimization module is specifically configured to: according to the predicted data of the key indicators, adopt an optimization algorithm to optimize the control parameters in the target optimization function, to obtain the optimized control parameters.

[0029] In a third aspect, the present application provides a parameter optimization apparatus, comprising: a processor, a memory, an input / output device and a bus; the memory stores computer instructions; when the processor executes the computer instructions in the memory, the memory stores computer instructions; and the processor, when executing the computer instructions in the memory, is configured to implement any one of the implementations of the first aspect.

[0030] In a fourth aspect, the embodiments of the present application provide a computer readable storage medium. The computer readable storage medium stores computer instructions; when the computer instructions run on a computer, the computer executes the method described in the possible implementation of the first aspect.

[0031] In a fifth aspect, the embodiments of the present application provide a computer program product. The computer program product comprises computer programs or instructions, and when the computer programs or instructions run on a computer, the computer executes the method described in the possible implementation of the first aspect. BRIEF DESCRIPTION OF DRAWINGS

[0032] FIG. 1 is a schematic diagram of a system framework provided by the present application;

[0033] FIG. 2 is a schematic diagram of a process parameter optimization method;

[0034] FIG. 3 is a schematic diagram of a parameter optimization method provided by the present application;

[0035] FIG. 4 is a schematic diagram of time series prediction based on frequency domain time series enhancement;

[0036] FIG. 5 is a comparison of output dosing frequencies under different modes, and the clear water layer height and overflow concentration curves corresponding to different dosing frequencies;

[0037] FIG. 6 is a comparison of output results with and without a penalty term, and the clear water layer height and overflow concentration curves corresponding to the output results;

[0038] FIG. 7 is a schematic diagram of the structure of a parameter optimization apparatus provided by the present application;

[0039] FIG. 8 is a schematic diagram of the structure of another parameter optimization apparatus provided by the present application. DETAILED DESCRIPTION

[0040] With reference to the drawings and the embodiments of the present application, the technical solutions in the embodiments of the present application will be described. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0041] The terms "first", "second", "third", "fourth" and the like in the description of the specification and claims of the present application and the above drawings, if any, are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "comprise" and "have" and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a list of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to such processes, methods, products or devices.

[0042] In the embodiments of the present application, the words "exemplary" or "for example" are used to mean serving as an example, instance, or illustration. Any embodiment or design presented as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Rather, the use of the words "exemplary" or "for example" is intended to present relevant concepts in a concrete manner.

[0043] In order to better understand the solutions of the embodiments of the present application, the related terms and concepts involved in the embodiments of the present application will be introduced first.

[0044] (1) Transformer

[0045] Transformers are a type of deep learning model based on attention mechanisms that have revolutionized the field of natural language processing (NLP). The core idea behind Transformers is to use self-attention and positional encoding to capture long-range dependencies within a sequence. The self-attention mechanism allows the model to consider other elements in the sequence while processing a certain element, which helps better understand the context. The architecture of a Transformer model mainly consists of two parts: an encoder and a decoder. The encoder is composed of multiple identical layers, each containing two sub-layers. The first sub-layer is a multi-head self-attention mechanism that allows the model to focus on multiple positions in the sequence simultaneously. The second sub-layer is a simple, positionally fully connected feed-forward neural network. The decoder is also composed of multiple identical layers, each containing three sub-layers. The first sub-layer is a multi-head self-attention layer that captures dependencies within the decoder. The second sub-layer is a multi-head attention layer that focuses on the output of the encoder. The third sub-layer is also a feed-forward neural network. Each sub-layer contains a residual connection and a layer normalization, which helps avoid the vanishing gradient problem in deep networks and speeds up the training process.

[0046] (2) Multi-Layer Perceptron

[0047] A multi-layer perceptron (MLP) is a type of feed-forward neural network composed of an input layer, one or more hidden layers, and an output layer. Each layer is composed of multiple neurons, and the layers are connected through fully connected connections. MLPs are fundamental models in neural networks and deep learning, suitable for various scenarios such as classification problems (text classification, image classification, or speech classification), regression problems (predictive analysis, etc.), time series prediction, and natural language processing.

[0048] (3) Pooling

[0049] Pooling is an important operation in Convolutional Neural Networks (CNNs) that reduces the spatial dimensions of data, thereby reducing computational load and parameter quantity, preventing overfitting, while preserving important feature information. Pooling operations are usually performed after convolutional layers, which helps to extract robust features and enhance the generalization ability of the model. Pooling can be divided into the following types: Max Pooling, Average Pooling, Global Max Pooling, and Global Average Pooling.

[0050] (4) Layer Normalization

[0051] The core idea of layer normalization (LN) is to normalize the activation values at each layer of the network, making the distribution of different features more stable and consistent.

[0052] (5) Overflow

[0053] The key indicator in the optimization of coal slime water dosing is the concentration of coal slime water on the surface of the coal washing pool.

[0054] (6) Height of the clear water layer

[0055] The key indicator in the optimization of coal slime water dosing is the height of the interface between the water layer and the mud layer in the coal washing pool.

[0056] The parameter optimization method provided by the present application can be applied to parameter optimization in industrial scenarios, such as intelligent dosing optimization of coal slime water in coal mines, process parameter optimization of refining furnaces, temperature control optimization of converters, dense medium coal separation optimization, etc. In addition, the time series prediction model in the embodiments of the present application can be applied to any time series prediction scenario or task, such as power load prediction, financial data prediction, roller abnormality detection, and other related time series tasks.

[0057] Some possible system architectures provided by the embodiments of the present application are introduced below.

[0058] Referring to FIG. 1, the embodiment of the present application provides a system architecture 100. As shown in the system architecture 100, an input module 110 can be used to obtain time series data. After the input module 110 obtains the time series data, a time series prediction module 120 can predict a key indicator according to the obtained time series data to obtain a prediction result of the key indicator, wherein the time series prediction module 120 includes a frequency domain multi-scale learning module 121, a time series information enhancement module 122, and a fusion module 123. Subsequently, a parameter optimization module 130 can construct a target optimization function according to the prediction result and a time series prediction model constructed in the time series prediction module 120, solve the target optimization function, and obtain an optimized control parameter.

[0059] As shown in FIG. 2, FIG. 2 is a flowchart of a process parameter optimization method, which includes two stages. In the first stage, a time series prediction model based on frequency domain time series enhancement is used to predict the future changes of a key process indicator in a process industry according to input time series data. In the second stage, a target optimization function can be constructed according to the prediction result of the key process indicator obtained in the first stage and a target value of the key process indicator. Subsequently, a trend-aware adaptive optimization algorithm can be used to optimize the process parameters in the target optimization function, and the process parameters, i.e., the input variables, are constantly updated, so that the prediction result of the key process indicator changes, and then the target optimization function changes. Finally, when the target optimization function takes an optimal value, the current process parameters are output, and the final optimized process parameters are obtained.

[0060] Currently, the optimization of process parameters in the process industry is mainly based on a PID controller or a standard evolutionary algorithm for parameter optimization. Since certain production requirements need to be met in an industrial scenario, directly using a standard evolutionary algorithm for parameter optimization may not be able to achieve the process target in the process industry.

[0061] In addition, due to poor industrial data quality and noise, the performance of classical long short-term memory (LSTM) and temporal convolutional network (TCN) models is poor, and it is difficult to meet the high-precision requirement. For complex transformer models such as Informer and Hidformer, it is difficult to apply them to small-sample industrial data.

[0062] To solve the current problem, an embodiment of the present application proposes a parameter optimization method, which can use a time series prediction model based on frequency domain time series enhancement to predict a key indicator, and optimize a control parameter according to the prediction result and a target optimization function to obtain an optimized control parameter. The target optimization function is determined based on the time series prediction model and a target value of the key indicator, so that the key indicator under the optimized control parameter can better meet the working condition standard.

[0063] Next, the method flow provided by the present application is introduced in combination with the foregoing system architecture.

[0064] Referring to FIG. 3, a flowchart of a parameter optimization method provided by the present application is shown as follows.

[0065] 301. Obtain time series data of a key indicator;

[0066] In an embodiment of the present application, first, time series data of a key indicator in a process industry can be obtained. The time series data is a data column recorded in time sequence for the key indicator. For example, in coal slime water dosing optimization, the key indicator can be overflow concentration or clear water layer height. The clear water layer height can be recorded in time sequence to obtain time series data of the clear water layer height. The key indicator can be used to optimize a control parameter, which can be a process parameter in the process industry. For example, in coal slime water dosing optimization, the control parameter can be dosing frequency.

[0067] 302. Use a time series prediction model to predict according to the time series data to obtain prediction data of the key indicator;

[0068] In the embodiments of the present application, the time series prediction model is a time series prediction model based on frequency domain time series enhancement (also referred to as a time series frequency enhancement model (TFAFormer)), which includes two branches. One branch can obtain frequency domain multi-scale features of time series data through a frequency domain multi-scale learning module (PFB) according to the time series data. The other branch can obtain time series features of the time series data through a time series enhancement module (TCB). Subsequently, the features obtained by the two branches can be fused, and the fused features can be predicted through a multi-layer perceptron (MLP) to obtain a prediction result (prediction data of the key indicator). The prediction result includes prediction data of the key indicator at different future time points, and can be used to indicate subsequent control parameter adjustment to make the key indicator as close as possible to the target value and make the target optimization function obtain the optimal value.

[0069] Specifically, the specific process of time series prediction based on frequency domain time series enhancement is shown in FIG. 4. As shown in FIG. 4, the time series data is N*L dimensional data, where N is the time dimension and L is the feature dimension. Based on the obtained N*L dimensional time series data, the time series data can be processed in parallel through K PFB modules and K TCB modules, respectively, where K is a positive integer. After obtaining the feature representation of the time series data through the PFB model and the time series features of the time series data through the TCB module, the feature representation and the time series features can be added to obtain fused features, and then the MLP is used for prediction to obtain the prediction data of the key indicator.

[0070] The time series data is subjected to frequency domain multi-scale feature extraction by the PFB module. First, the time series data can be subjected to fast Fourier transform, converting the time domain signal of the time series data to the frequency domain to obtain the frequency domain features of the time series data. Then, the obtained frequency domain features can be subjected to multi-scale pooling for dimension reduction processing of the frequency domain features, thereby reducing the calculation amount in the prediction process. For example, the frequency domain features can be subjected to maximum pooling or average pooling, and the specific pooling manner is not limited. After multi-scale pooling, multi-scale feature extraction can be performed through a full connection (FC) layer to obtain multi-scale features. Then, the multi-scale features can be up-sampled to restore the dimension of the initial time series data, and the multi-scale features can be converted into multi-scale time domain features through inverse Fourier transform. The multi-scale time domain features and the time series features of the time series data are added through layer normalization, and then the feature representation of the time series data is obtained. After obtaining the feature representation of the time series data, the feature representation can be improved and refined through a feature extraction module. The feature extraction module includes a multi-layer perception, a residual connection and layer normalization.

[0071] The time series data is subjected to time series feature extraction by the TCB module. First, the time series data can be subjected to time series convolution to extract the time series features of the time series data. Then, the time series features output by each network layer are normalized through layer normalization to eliminate internal covariate shift in the model training process. In addition, similar to the PFB module, after obtaining the time series features of the time series data, the time series features can be refined through a feature extraction module to make the extracted time series features more accurate, and to make the prediction result obtained based on the time series features more accurate.

[0072] In the embodiments of the present application, the frequency domain multi-scale learning module can be used to extract multi-scale information of the time series data in the frequency domain, so that the extracted features are more abundant. In addition, multi-scale pooling is performed on the frequency domain features, which helps to filter noise and makes the time series prediction model more robust.

[0073] 303、According to the time series prediction model and the target value of the key indicator, a target optimization function is determined;

[0074] Before determining the target optimization function, the target value of the key indicator can also be obtained. The key indicator can be one or more important indicators for evaluating the time series task. For example, in the coal slime water dosing optimization process, the key indicators can be overflow and / or clear water layer height. Therefore, the target value corresponding to the key indicator can also be one or more. The selection of the key indicator and the target value can be set according to the actual situation, and the specific implementation is not limited herein.

[0075] Optionally, a first loss function in the target optimization function can be determined according to a time series prediction model and a target value of the key indicator, the first loss function can be calculated according to a second loss function and a coefficient corresponding to the second loss function, and the second loss function can represent a difference between a current value of the key indicator and the target value corresponding to the key indicator; after the first loss function is obtained, the target optimization function can be determined according to a weight corresponding to the first loss function.

[0076] The weight can be self-defined according to different prediction results, and different weights can be set according to prediction times from near to far. For example, the prediction results include the key indicator change trend at three time points of 5 minutes, 10 minutes and 15 minutes in the future, and the corresponding weights can be set to 0.2, 0.5 and 0.7 respectively. The calculation formula of the first loss function can be where w i is the weight value of the first loss function at the i th time point, is used to represent the value of the key indicator corresponding to the control parameter x i . Therefore, the expression of the target optimization function can be The target optimization function can represent the difference between the value of the key indicator corresponding to different prediction times and the target value.

[0077] In the embodiments of the present application, different weights can be set for the first loss function according to different prediction times to adapt to the actual change trend of the key indicator, so that the final optimization result can better meet the production requirements.

[0078] The parameter optimization in the industrial scene is usually optimized by using an evolutionary algorithm, but since the standard optimization algorithm does not combine the actual working conditions, it cannot adapt to the actual working conditions. Therefore, in the embodiments of the present application, the difference between the key indicator and the target value can be combined to optimize the control parameter, so that the key indicator can be as close to the target value as possible, thereby adapting to the actual working conditions.

[0079] Optionally, the second loss function can be determined by judging whether the key indicator under the current control parameter x i meets the preset requirement, and the determination result can be represented by a value function. When the key indicator does not meet the preset requirement, the second loss function needs to be quickly adjusted, and therefore the second loss function can be multiplied by a larger coefficient (a first coefficient) to make the first loss function change more greatly and the optimal value can be obtained more quickly. It should be understood that when the key indicator does not meet the preset requirement, the second loss function can be multiplied by a smaller coefficient (a second coefficient), and the second coefficient can be the inverse of the first coefficient.

[0080] The value function can be a sign function, When the preset requirement is that the key indicator f(x,θ)>t1, the coefficient expression can be μ-sign(f(x,θ)-t1). When the key indicator does not meet the preset requirement, the expression of the first loss function can be:

[0081] When the key indicators meet the preset requirements, the expression of the first loss function can be:

[0082] In addition, corresponding target values ​​can be set for different key indicators. For example, in the optimization of coal slime water dosing, the key indicators can be overflow and clear water layer height. The two key indicators correspond to two target values. In this case, the expression of the first loss function can be:

[0083] Among them, α is the weight used to balance the two objectives.

[0084] In addition, in order to reduce the sudden change of the output control parameters, a penalty term can be added to the first loss function to make the output control parameters more stable. Therefore, the expression of the first loss function can also be:

[0085] Where X is the value of the current control parameter. The expression of the first loss function can be determined according to the selection of key indicators and the setting of penalty terms, which are not limited here. The penalty term can be the square of the difference between the control parameter and the current control parameter, or the absolute value of the difference between the control parameter and the current control parameter, which are not limited here. For example, in the optimization of coal slurry dosing, the control parameter is the dosing frequency, so x i is the dosing frequency, X is the current dosing frequency, and the dosing frequency x can be continuously optimized. i , so that the objective optimization function obtains the minimum value, that is, the key indicators overflow and clear water layer height are as close to the corresponding target values ​​as possible.

[0086] 304. Optimize the control parameters in the target optimization function according to the predicted data of the key indicators to obtain the optimized control parameters.

[0087] Among them, the predicted data of the key indicators may include the values ​​corresponding to the key indicators at different future moments, which can represent the changing trend of the key indicators. Therefore, the control parameters can be adjusted according to the predicted data to make the key indicators closer to the target values.

[0088] Exemplarily, when the value of the key indicator is less than the target value, if increasing the value of the control parameter makes the value of the key indicator smaller, the value of the control parameter can be decreased according to the change trend of the key indicator; if increasing the value of the control parameter makes the value of the key indicator larger, the value of the control parameter can be continuously increased, so as to realize the optimization adjustment of the control parameter to make the value of the key indicator as close as possible to the target value.

[0089] Optionally, the control parameter in the target optimization function can be adjusted by using an evolutionary algorithm according to the prediction result, so that the target optimization function obtains an optimal value, that is, the difference between the value of the key indicator obtained by prediction and the target value is minimized, and then the control parameter at this time, that is, the optimized control parameter, is output. The evolutionary algorithm can be a genetic algorithm, and can also be a particle swarm algorithm, which is not limited here.

[0090] In the embodiment of the application, the prediction result of the time series prediction model on the time series data can be used to represent the change trend of the key indicator in the time series data, and therefore, the control parameter in the target optimization function can be optimized and adjusted according to the prediction result, so that the target optimization function obtains an optimal value, that is, the key indicator is as close as possible to the target value, to adapt to the actual working condition target.

[0091] At present, some existing prediction methods are, for example: using an informer model, a Dlinear model, an LGBM (light gradient boosting machine) model, a TCN model or a Hidformer model to predict the time series data.

[0092] Based on the foregoing existing prediction methods, two key indicators, overflow and clear water layer height, in coal slime water dosing optimization are exemplarily predicted, and the existing prediction methods are compared with the prediction method provided in the embodiment of the application.

[0093] The comparison result can be shown in Table 1.

[0094] Table 1

[0095] As can be seen from Table 1, the time series prediction model provided in the embodiment of the application exceeds the advanced method in the industry in the prediction of overflow and clear water layer height. In the prediction of overflow, the MAE of the application is 0.0949, which is less than 0.1187 of Hidformer and 0.2189 of Informer. In the prediction of clear water layer height, the MAE of the application is 0.0851, which is less than 0.1484 of Hidformer and 0.1512 of Dlinear.

[0096] The prediction model proposed in the embodiments of the present application is compared with a single branch PFB, a single branch TCB, a two-branch PFB+TCB serial mode, a frequency domain single scale learning (TFAFormer-SS), and a frequency domain multi-scale learning (TFAFormer-MS), wherein the PFB branch adopts frequency domain multi-scale learning, and TFAFormer-SS and TFAFormer-MS both adopt the PFB and TCB parallel mode. The specific comparison results can be shown in Table 2.

[0097] Table 2

[0098] As shown in Table 2, the MAE and RMSE of the single branch and the two-branch serial mode are both greater than the two-branch parallel mode, so it can be seen that the training effect of the dual-branch parallel structure proposed in the present application is better; the MAE of the frequency domain single scale learning is 0.1448, and the RMSE is 0.1876, while the MAE of the frequency domain multi-scale learning is 0.0949, and the RMSE is 0.1345, so the effect of the frequency domain multi-scale learning is better than that of the frequency domain single scale learning.

[0099] In the embodiments of the present application, in the optimization of the control parameters, the first loss function at different prediction times can be weighted, so that the target optimization function can be dynamically adjusted according to the weights of different prediction times to realize trend perception. And according to the difference between the current key indicators and the target value, the loss function can be dynamically adjusted by different coefficients to realize dynamic adaptation to the working conditions. For example, taking coal slime water dosing optimization as an example, various different situations are compared and analyzed. The comparison results can be shown in Table 3.

[0100] Table 3

[0101] As shown in Table 3, in the coal slime water dosing optimization, the dosing frequency is inversely proportional to the clear water layer height and is proportional to the overflow, and the dosing effect is the best. Therefore, the linear regression coefficient is used as an evaluation index, and the larger the absolute value of the linear regression coefficient is, the better. After increasing the trend perception, the linear regression coefficient of the clear water layer decreases from-5.139 to-6.763, and the linear regression coefficient of the overflow increases from 1.875 to 3.841, proving that the dosing frequency response increases.

[0102] After increasing the adaptive function, the linear regression coefficient of the clear water layer decreases to-7.412, and the linear regression coefficient of the overflow increases to 4.910, proving that the dosing frequency responds more to the working conditions.

[0103] As can be seen from the output dosing frequency curve shown in FIG. 5, the scheme is faster in responding to index changes and is more stable with less jumps. As can be seen from the clear water layer height and overflow concentration curves in FIG. 5, the dosing frequency is negatively correlated with the clear water layer height and positively correlated with the overflow concentration. The smaller the clear water layer height and the greater the overflow concentration, the greater the expected dosing frequency.

[0104] Exemplarily, in the coal slime water dosing optimization, the embodiment of the application can increase a penalty term in the loss function to make the output result more stable. The comparison between the penalty term and the non-penalty term can be as shown in FIG. 6.

[0105] The foregoing describes the method flow provided by the application, and the following describes the device provided by the application based on the foregoing method flow.

[0106] Referring to FIG. 7, the structure schematic diagram of a parameter optimization device provided by the application is as follows.

[0107] The acquisition module 701 is configured to acquire time series data of a key index, the key index being used for optimizing a control parameter;

[0108] The prediction module 702 is configured to perform prediction according to the time series data by using a time series prediction model to obtain prediction data of the key index.

[0109] The determination module 703 is configured to determine a target optimization function according to the time series prediction model and a target value of the key index, the target optimization function representing a difference between the prediction data of the key index and the target value.

[0110] The optimization module 704 is configured to optimize the control parameter in the target optimization function according to the prediction data of the key index to obtain an optimized control parameter.

[0111] In a possible implementation, the foregoing prediction module 702 is specifically configured to: perform frequency domain multi-scale learning on the time series data to obtain a feature representation of the time series data, the feature representation including multi-scale features of the time series data; extract time series information of the time series data by using time series convolution to obtain time series features of the time series data; and perform prediction by using a time series prediction model according to the feature representation and the time series features to obtain the prediction data of the key index.

[0112] In a possible implementation, the foregoing prediction module 702 is specifically configured to: perform prediction by using a multi-layer perception according to the feature representation and the time series features to obtain the prediction data of the key index.

[0113] In a possible implementation, the prediction module 702 is specifically configured to: perform fast Fourier transform on the time series data to obtain frequency domain features of the time series data; perform multi-scale pooling on the frequency domain features to obtain pooled frequency domain features; perform multi-scale feature extraction on the pooled frequency domain features to obtain multi-scale features; and perform inverse fast Fourier transform on the multi-scale features to obtain the feature representation.

[0114] In a possible implementation, the determination module 703 is specifically configured to: determine, according to the time series prediction model and the target value of the key indicator, a first loss function in the target optimization function, the first loss function representing a difference between the current value and the target value of the key indicator; and obtain the target optimization function according to the first loss function and a weight of the first loss function.

[0115] In a possible implementation, the determination module 703 is specifically configured to: determine, according to the time series prediction model and the target value, whether the key indicator meets a preset requirement; if the preset requirement is not met, determine the first loss function according to a first coefficient and a second loss function, the second loss function being a difference between the current value and the target value of the key indicator; and if the preset requirement is met, determine the first loss function according to a second coefficient and the second loss function, the second coefficient being an inverse of the first coefficient.

[0116] In a possible implementation, the optimization module 704 is specifically configured to: optimize, according to the predicted data of the key indicator, a control parameter in the target optimization function by using an optimization algorithm to obtain an optimized control parameter.

[0117] Referring to FIG. 8, another structural schematic diagram of a parameter optimization apparatus is provided, and the parameter optimization apparatus is described as follows.

[0118] The parameter optimization apparatus can include a processor 801 and a memory 802. The processor 801 and the memory 802 are connected through a line. The memory 802 stores program instructions and data.

[0119] The memory 802 stores program instructions and data corresponding to the steps in FIG. 3.

[0120] The processor 801 is configured to execute the method steps performed by the parameter optimization apparatus shown in FIG. 3.

[0121] Optionally, the parameter optimization apparatus can further include a transceiver 803 configured to receive or send data.

[0122] In the embodiments of the present application, a computer readable storage medium is also provided, and the computer readable storage medium stores a program. When the program is run on a computer, the computer is caused to execute the steps in the method described in the foregoing embodiments shown in FIG. 3.

[0123] The embodiment of the present application also provides a computer program product comprising the computer program product, which, when running on a computer, causes the computer to execute the method steps described in the foregoing embodiment shown in Fig. 3.

[0124] In addition, it should be noted that the above-described device embodiments are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment.

[0125] Through the description of the above embodiments, those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0126] In the several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented by other ways. For example, the above-described device embodiments are merely illustrative, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0127] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0128] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of software functional unit.

[0129] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, read-only memory), a random access memory (RAM, random access memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0130] Finally, it should be noted that: the above, only for the specific embodiments of the present application, but the protection scope of the present application is not limited to this, any person skilled in the art within the technical scope disclosed in the present application, can easily think of changes or replacements, should be covered within the protection scope of the present application.

Claims

1. A parameter optimization method, characterized in that: include: Obtaining time series data of key indicators, wherein the key indicators are used to optimize control parameters; Based on the time series data, a time series prediction model is used to perform prediction to obtain predicted data of the key indicators; Determining a target optimization function based on the time series prediction model and the target value of the key indicator, wherein the target optimization function represents the difference between the predicted data of the key indicator and the target value; According to the predicted data of the key indicators, the control parameters in the target optimization function are optimized to obtain optimized control parameters.

2. The method according to claim 1, characterized in that The step of using a time series prediction model to perform prediction based on the time series data to obtain prediction data of the key indicators includes: Performing frequency domain multi-scale learning on the time series data to obtain a feature representation of the time series data, wherein the feature representation includes multi-scale features of the time series data; Extracting time series information of the time series data using time series convolution to obtain time series features of the time series data; According to the feature representation and the time series feature, the time series prediction model is used to perform prediction to obtain prediction data of the key indicator.

3. The method according to claim 2, characterized in that The step of performing prediction based on the feature representation and the time series feature using the time series prediction model to obtain prediction data of the key indicator includes: According to the feature representation and the time series feature, a multi-layer perceptron is used to perform prediction to obtain the predicted data of the key indicator.

4. The method according to any one of claims 2 or 3, characterized in that The performing frequency domain multi-scale learning on the time series data to obtain a feature representation of the time series data includes: Performing a fast Fourier transform on the time series data to obtain frequency domain features of the time series data; Performing multi-scale pooling on the frequency domain features to obtain pooled frequency domain features; Performing multi-scale feature extraction on the pooled frequency domain features to obtain multi-scale features; The feature representation is obtained by performing an inverse fast Fourier transform on the multi-scale features.

5. The method according to any one of claims 1 to 4, characterized in that Determining a target optimization function according to the time series prediction model and the target value of the key indicator includes: Determining a first loss function in the objective optimization function according to the time series prediction model and the target value of the key indicator, wherein the first loss function represents a difference between a current value of the key indicator and the target value; The target optimization function is obtained according to the first loss function and the weight of the first loss function.

6. The method according to claim 5, characterized in that Determining the first loss function in the objective optimization function according to the time series prediction model and the target value of the key indicator includes: Determining whether the key indicator meets preset requirements based on the time series prediction model and the target value; If the preset requirement is not met, determining the first loss function according to the first coefficient and a second loss function, where the second loss function is the difference between the current value of the key indicator and the target value; If the preset requirement is met, the first loss function is determined according to the second coefficient and the second loss function, where the second coefficient is the reciprocal of the first coefficient.

7. The method according to any one of claims 1 to 6, characterized in that Optimizing the control parameters in the target optimization function according to the predicted data of the key indicators to obtain the optimized control parameters includes: According to the predicted data of the key indicators, an optimization algorithm is used to optimize the control parameters in the target optimization function to obtain the optimized control parameters.

8. A parameter optimization device, characterized in that: include: An acquisition module, configured to acquire time series data of key indicators, wherein the key indicators are used to optimize control parameters; A prediction module, configured to perform prediction based on the time series data using a time series prediction model to obtain predicted data of the key indicators; A determination module, configured to determine a target optimization function based on the time series prediction model and the target value of the key indicator; The optimization module is used to optimize the control parameters in the target optimization function according to the prediction data of the key indicators to obtain the optimized control parameters.

9. The device according to claim 8, characterized in that The prediction module is specifically used to: Performing frequency domain multi-scale learning on the time series data to obtain a feature representation of the time series data, wherein the feature representation includes multi-scale features of the time series data; Extracting time series information of the time series data using time series convolution to obtain time series features of the time series data; According to the feature representation and the time series feature, the time series prediction model is used to perform prediction to obtain prediction data of the key indicator.

10. The device according to claim 9, characterized in that The prediction module is specifically used to: According to the feature representation and the time series feature, a multi-layer perceptron is used to perform prediction to obtain the predicted data of the key indicator.

11. The device according to any one of claims 9 or 10, characterized in that The prediction module is specifically used to: Performing a fast Fourier transform on the time series data to obtain frequency domain features of the time series data; Performing multi-scale pooling on the frequency domain features to obtain pooled frequency domain features; Performing multi-scale feature extraction on the pooled frequency domain features to obtain multi-scale features; The feature representation is obtained by performing an inverse fast Fourier transform on the multi-scale features.

12. The device according to any one of claims 9 to 11, characterized in that The determining module is specifically configured to: Determining a first loss function in the objective optimization function according to the time series prediction model and the target value of the key indicator, wherein the first loss function represents a difference between a current value of the key indicator and the target value; The target optimization function is obtained according to the first loss function and the weight of the first loss function.

13. The device according to claim 12, characterized in that The determining module is specifically configured to: Determining whether the key indicator meets preset requirements based on the time series prediction model and the target value; If the preset requirement is not met, determining the first loss function according to the first coefficient and a second loss function, where the second loss function is the difference between the current value of the key indicator and the target value; If the preset requirement is met, the first loss function is determined according to the second coefficient and the second loss function, where the second coefficient is the reciprocal of the first coefficient.

14. The device according to any one of claims 8 to 13, characterized in that The optimization module is specifically used to: According to the predicted data of the key indicators, an optimization algorithm is used to optimize the control parameters in the target optimization function to obtain the optimized control parameters.

15. A parameter optimization device, characterized in that: include: A data processing device, comprising: a processor and a memory, the processor being coupled to the memory; The memory is used to store programs; The processor is configured to execute the program in the memory so as to perform the method according to any one of claims 1 to 7. 16 . A computer-readable storage medium comprising instructions, which, when executed on a computer, enable the computer to perform the method according to claim 1 .

17. A computer program product comprising instructions which, when run on a computer, cause the computer to perform the method according to any one of claims 1 to 7.

Images