Chemical disaster prediction method based on sequence feature decomposition algorithm
By combining sequence feature decomposition algorithm and dendritic learning model, time series data of chemical production is decomposed and processed, the problem of insufficient chemical disaster prediction accuracy in the existing technology is solved, and more accurate and timely disaster warning is achieved, and more accurate and timely disaster warning is improved. Chemical safety management level.
Patent Information
- Application Number
- CN202410654394.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-24
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-05-24
AI Technical Summary
When existing chemical disaster prediction technologies process complex and nonlinear time series data, the prediction accuracy is insufficient, it is difficult to capture nonlinear features in the data, and the calculation load is too large, making it difficult to achieve fast and effective prediction.
The time series data of chemical production is processed using sequence feature decomposition algorithm (SFDL), decomposed into seasonal, trend-period and residual components, and combined with dendritic learning models, pattern recognition and prediction are performed through multi-layer neural networks. , enhance the nonlinear processing capability of the model and its robustness to abnormal data.
Improve the accuracy and timeliness of chemical disaster prediction, enable disaster warnings earlier, reduce the complexity of model construction and dependence on professional knowledge, and achieve rapid response and dynamic updates, suitable for the chemical industry and other time-consuming The field of sequence prediction.
Smart Images

Figure CN118445591B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of chemical disaster prevention decision-making, and in particular to a chemical disaster prediction method based on a sequence feature decomposition algorithm. Background Art
[0002] Safety management in chemical production environments has always been a major challenge in the field of industrial safety. The production parks of chemical plants are usually equipped with complex transportation pipeline systems for transporting various chemical raw materials. If these pipelines are not managed properly, they are very likely to cause chemical reactions due to improper intersection of raw materials, leading to serious safety accidents. Such accidents not only threaten the safety of personnel, but may also cause huge economic losses and trigger chain reactions within the enterprise.
[0003] Although a variety of technologies have been developed to predict and prevent such disasters, they each have their limitations. For example, some invented autoregressive techniques use historical observation data to predict future events. Although this technology performs well in processing regular data, it is difficult to capture its nonlinear characteristics when faced with irregular or chaotic data, resulting in insufficient prediction accuracy. In addition, although the numerical prediction method based on physical principles provides a systematic solution in theory, in practical applications, due to the complexity of chemical processes, the variability of participating parameters, and potential continuity problems between data, the computational load of this method is too large, making it difficult to achieve fast and effective predictions.
[0004] Therefore, the existing chemical disaster prediction technology still has many shortcomings, especially in terms of data processing and the accuracy of the prediction model. These limitations indicate that there is still a lot of room for improvement in the algorithms and models in the field of chemical disaster prediction, and it is urgent to develop new prediction methods to more accurately and effectively process complex and nonlinear time series data, thereby improving the overall effectiveness of chemical safety management. Summary of the invention
[0005] The purpose of the present invention is to provide a chemical disaster prediction method based on a sequence feature decomposition algorithm to address the defects of the prior art, with the purpose of effectively predicting disaster events that may occur in the chemical production process to improve the efficiency and accuracy of chemical safety management. By decomposing the time series data related to chemical production, the method can identify and process the seasonal and trend components in the data, and combined with the dendritic learning model, further enhance the model's nonlinear processing ability and robustness to abnormal data, thereby providing more accurate and timely disaster warnings.
[0006] To achieve the above object, the present invention provides a chemical disaster prediction method based on a sequence feature decomposition algorithm, and the specific steps are as follows:
[0007] S1: The SFDL decomposition algorithm is used to process the time series data of chemical production, and the original time series data is decomposed into three main components: seasonal component, trend-cycle component and residual component, each of which is processed and analyzed by a specific mathematical model to improve the accuracy of prediction and the efficiency of data processing;
[0008] S2: First, the seasonal component is treated with Fourier synchronous compression transform FSST to make the seasonal component more prominent, and the component data is denoised to make the data more regular. Then, the seasonal component and trend-cycle component obtained by decomposition are further smoothed, and the local weighted regression algorithm is used to optimize and adjust these two components to enhance the adaptive ability to trend and seasonal changes in time series.
[0009] S3: Combined with the dendritic learning model, a multi-layer neural network model is constructed to perform pattern recognition on the decomposed and smoothed data, including the synaptic layer for signal transmission simulation, the dendritic layer for signal processing, the membrane layer for signal integration and output, and the somatic layer and axon layer for the final output prediction results;
[0010] S4: Apply the data obtained from the above processing and analysis to predict chemical disasters. By comparing the actual production data with the model prediction results, adjust the model parameters and prediction strategies to achieve early warning and accurate prediction of possible disaster events in chemical production.
[0011] As a preferred technical solution of the present invention, the SFDL decomposition algorithm described in step S1 specifically comprises the following steps for sequence decomposition:
[0012] S1-1: Decompose the original time series data into three subsequences. The specific calculation formula is:
[0013] Y k =T k +S k +R k (1)
[0014] Among them, Y k Indicates the time series data used; S k represents the seasonal term; T k represents the trend term-cycle term; R k represents the residual term; k is 1 to N, representing the length of the sequence;
[0015] S1-2: SFDL decomposition algorithm uses independent variable x i and the dependent variable y i Regression curve of two parameters Where i = 1, 2, ..., n, The fitted value of is the smoothness of y along the scale of the independent variable;
[0016] S1-3: From x i Select the point r closest to x and assign it a neighborhood weight, and make the following judgments:
[0017] If r≤n, from {x i} selects the r points closest to x, and then each point will be assigned a specific local weight according to its distance from x, which is calculated as:
[0018]
[0019]
[0020] Among them, λ r (x) represents the point x farthest from r i distance, w is the cubic weight function, v i (x) is x i The local weight of , t represents the normalized distance;
[0021] If r>n, λ r The definition of (x) becomes the following formula:
[0022]
[0023] Among them, the x closest to x i The weight is the largest; when x i As we move away from x, the weight decreases and eventually drops to 0; finally at point (x i ,y i ) is calculated The polynomial function of x is obtained by local fitting value and
[0024] As a preferred technical solution of the present invention, the specific operation steps of step S2 are as follows:
[0025] S2-1: Processing time series data in an iterative detrending loop, which consists of two recursive loops: an inner loop and an outer loop, where the inner loop is responsible for updating and smoothing the seasonal component of each generated subsequence; and the outer loop is responsible for updating and smoothing the trend component.
[0026] S2-2: In each iteration, a robustness weight is calculated based on the estimated residual component. This weight is used to reduce the impact of outlier observations in subsequent calculations and improve the model's ability to handle abnormal data.
[0027] S2-3: Through the above iterative processing, the expression of seasonal and trend components is gradually refined to adapt to the complexity and dynamic changes of time series data in the chemical production process.
[0028] As a preferred technical solution of the present invention, the structure of the dendritic learning model in step S3 is as follows: a synaptic layer, which uses a Sigmoid function as a processing tool to simulate signal transmission between neurons, receives input signals and calculates outputs based on connection weights and thresholds to simulate the excitation or inhibition state of neurons; a dendritic layer, which includes a multiplication unit for receiving input signals from the synaptic layer, and processes the combined effects of multiple inputs to enhance the information processing capability and nonlinear characteristics of the model; a membrane layer, as the third layer, is used to integrate output signals from the dendritic layer, and realize signal aggregation through an addition unit to form a response to the complex input of the previous layer; a somatic layer, which uses a Tanh function to process the output of the membrane layer, and the axon layer finally generates the output signal of the model, which represents the predicted possibility or risk level of chemical hazards, wherein the synaptic layer is the Synaptic layer, the dendritic layer is the Dendritic layer, the membrane layer is the Membrane layer, the somatic layer is the Soma layer, and the axon layer is the Axon layer.
[0029] As a preferred technical solution of the present invention, the function and structure of the synaptic layer are as follows: the synaptic layer is mainly responsible for simulating signal transmission between neurons, and the original signal is transmitted from a presynaptic neuron to another neuron through the synaptic cleft to achieve effective transmission of information; in this layer, the Sigmoid function is used to simulate the change of the postsynaptic potential, which is used to determine whether the synapse is excitatory or inhibitory, wherein, for the i-th synaptic input to the j-th synaptic layer, the expression of the Sigmoid function is as follows:
[0030]
[0031] Where i = 1, 2, ..., N; j = 1, 2, ..., M; x i ∈[0,1] represents the i-th input of the synapse; Y ij is the output of the jth synaptic layer; k is a positive constant; weight w ij and threshold θ ij are required connection parameters.
[0032] As a preferred technical solution of the present invention, the function and structure of the dendritic layer are as follows: the dendritic layer is responsible for processing the output signal from the synaptic layer. This layer processes the combination of multiple input signals by implementing a multiplication function, thereby enhancing the model's ability to capture the nonlinear characteristics of the signal. In this layer, the multiplication operator ∏ is used to achieve signal integration. The output of each dendritic branch is formed by the product of all the signals input to the branch, which is specifically expressed as the following formula:
[0033]
[0034] Among them, Z j Represented as the jth dendritic branch.
[0035] As a preferred technical solution of the present invention, the function and structure of the membrane layer are as follows: the membrane layer is mainly responsible for integrating all output signals from the dendritic layer. The layer uses an addition function to process the output signal of each dendritic branch, and standardizes the signal by logical or operation to form a unified output signal; the output V of the membrane layer is obtained by summing the outputs of all dendritic branches, and the specific expression is:
[0036]
[0037] Where V represents the output of the membrane layer.
[0038] As a preferred technical solution of the present invention, the function and structure of the somatic layer are as follows: the somatic layer uses the Tanh function to process the output signal from the membrane layer, which is used to simulate the activation and response process of neurons. When the output signal exceeds the set threshold, the somatic layer generates an activation signal; the output O of the somatic layer is expressed as:
[0039]
[0040] Among them, v represents the membrane layer, s represents the somatic layer, and the parameter k s is a positive number, the threshold θ s Changes from 0 to 1.
[0041] As a preferred technical solution of the present invention, the axon layer will discretely determine the weight of each input corresponding to the synaptic layer through the processed somatic layer output, that is:
[0042]
[0043]
[0044]
[0045] Among them, A represents the axon layer, k represents a normal number, std represents the process of calculating the standard deviation, and num represents the number of neurons.
[0046] As a preferred technical solution of the present invention, the connection types and constraints of the synapses in the synaptic layer in step S3 are as follows:
[0047] Case 1: Constant 0 connections, at w ij <0<θ ij Or 0<w ij <θij Under the condition, input x i When it changes from 0 to 1, the output is y ij is approximately 0;
[0048] Case 2: Constant 1 connection, at θ ij <0<w ij Or θ ij <w ij <0, input x i When it changes from 0 to 1, the output is y ij About 1;
[0049] Case 3: Excitatory connection, 0<θ ij <w ij Under the condition, input x i When it changes from 0 to 1, the output is Y ij There is a positive correlation;
[0050] Case 4: Inhibitory connection, in w ij <θ ij <0, input x i When changing from 0 to 1, the output is inversely proportional to the input.
[0051] Compared with the prior art, the present invention has the following beneficial effects:
[0052] High data adaptability and accuracy: By using the SFDLM algorithm, the present invention can more accurately decompose the seasonal and trend components in time series data. This fine decomposition helps the model capture potential abnormal patterns and trend changes, thereby providing earlier warnings of possible disasters in the chemical production process.
[0053] Enhanced nonlinear processing capability: Combining the decomposition function of SFDL and the learning mechanism of DLM, the present invention effectively handles the nonlinear problems of time series data. By using sigmoid function, Tanh function and multiplication unit in the DLM model, complex biological neural networks can be effectively simulated, improving the model's understanding and response capabilities to unconventional data dynamics.
[0054] No need to preset mathematical models: Unlike traditional prediction methods, the present invention does not rely on pre-built specific mathematical models, but directly learns and extracts necessary features from the data, which reduces the complexity of model construction and dependence on professional knowledge.
[0055] Real-time and dynamic update capabilities: Through continuous iterative cycles and real-time data processing, the present invention can dynamically update the prediction results and promptly reflect any important changes in the production process, thereby achieving rapid response and disaster prevention.
[0056] Wide application potential: Due to the high efficiency and accuracy of the invention in processing data, it is not only suitable for the chemical industry, but can also be extended to other fields that require time series prediction and disaster prevention, such as environmental monitoring, energy management, etc.
[0057] Improve safety management level: The present invention can provide chemical enterprises with a more scientific and systematic safety early warning mechanism, greatly improve the safety management level and disaster response capabilities of enterprises, and effectively protect the safety of personnel and property. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 A flow chart of a chemical disaster prediction method based on a sequence feature decomposition algorithm provided by the present invention;
[0059] Figure 2 The present invention provides prediction results of seven algorithms in the USA data set according to the embodiments;
[0060] Figure 3 The present invention provides prediction results of seven algorithms in the EU dataset according to the embodiments;
[0061] Figure 4 The present invention provides seven algorithms for predicting results on the CHN dataset. DETAILED DESCRIPTION
[0062] The present invention is further described below in conjunction with the accompanying drawings and examples. However, the present invention can be implemented in many different ways and should not be construed as being limited to the embodiments shown; on the contrary, these embodiments provide those skilled in the art with implementation methods that meet applicable legal requirements.
[0063] Example: Figure 1 As shown, the present invention discloses a chemical disaster prediction method based on a sequence feature decomposition algorithm, and the specific steps are as follows:
[0064] S1: The SFDL decomposition algorithm is used to process the time series data of chemical production, and the original time series data is decomposed into three main components: seasonal component, trend-cycle component and residual component, each of which is processed and analyzed by a specific mathematical model to improve the accuracy of prediction and the efficiency of data processing;
[0065] The SFDL decomposition algorithm performs sequence decomposition in the following steps:
[0066] S1-1: Decompose the original time series data into three subsequences. The specific calculation formula is
[0067] for:
[0068] Y k =T k +Sk +R k (1)
[0069] Among them, Y k Indicates the time series data used; S k represents the seasonal term; T k represents the trend term-cycle term; R k represents the residual term; k is 1 to N, representing the length of the sequence;
[0070] S1-2: SFDL decomposition algorithm uses independent variable x i and the dependent variable y i Regression curve of two parameters Where i = 1, 2, ..., n, The fitted value of is the smoothness of y along the scale of the independent variable;
[0071] S1-3: From x i Select the point r closest to x and assign it a neighborhood weight, and make the following judgments:
[0072] If r≤n, from {x i} selects the r points closest to x, and then each point will be assigned a specific local weight according to its distance from x, which is calculated as:
[0073]
[0074]
[0075] Among them, λ r (x) represents the point x farthest from r i distance, w is the cubic weight function, v i (x) is x i The local weight of , t represents the normalized distance;
[0076] If r>n, λ r The definition of (x) becomes the following formula:
[0077]
[0078] Among them, the x closest to x i The weight is the largest; when x i As we move away from x, the weight decreases and eventually drops to 0; finally at point (x i ,y i ) is calculated The polynomial function of x is obtained by local fitting value and
[0079] S2: Further smooth the seasonal component and trend-cycle component obtained by decomposition, and use the local weighted regression algorithm to optimize and adjust these two components to enhance the adaptability to trend and seasonal changes in time series;
[0080] The specific steps include:
[0081] S2-1: Processing time series data in an iterative detrending loop, which consists of two recursive loops: an inner loop and an outer loop, where the inner loop is responsible for updating and smoothing the seasonal component of each generated subsequence; and the outer loop is responsible for updating and smoothing the trend component.
[0082] S2-2: In each iteration, a robustness weight is calculated based on the estimated residual component. This weight is used to reduce the impact of outlier observations in subsequent calculations and improve the model's ability to handle abnormal data.
[0083] S2-3: Through the above iterative processing, the expression of seasonal and trend components is gradually refined to adapt to the complexity and dynamic changes of time series data in the chemical production process;
[0084] S3: Combined with the dendritic learning model, a multi-layer neural network model is constructed to perform pattern recognition on the decomposed and smoothed data, including the synaptic layer for signal transmission simulation, the dendritic layer for signal processing, the membrane layer for signal integration and output, and the somatic layer and axon layer for the final output prediction results;
[0085] The structure of the dendritic learning model is as follows: the synaptic layer, which uses the Sigmoid function as a processing tool to simulate the signal transmission between neurons, receives input signals and calculates outputs based on connection weights and thresholds to simulate the excitation or inhibition state of neurons; the dendritic layer, which contains multiplication units for receiving input signals from the synaptic layer, processes the combined effects of multiple inputs to enhance the information processing capabilities and nonlinear characteristics of the model; the membrane layer, as the third layer, is used to integrate the output signals from the dendritic layer, and realizes signal aggregation through the addition unit to form a response to the complex input of the previous layer; the somatic layer, which uses the Tanh function to process the output of the membrane layer, and the axon layer processes the output of the somatic layer to finally generate the output signal of the model, which represents the predicted possibility or risk level of chemical hazards, among which the synaptic layer is the Synaptic layer, the dendritic layer is the Dendritic layer, the membrane layer is the Membrane layer, the somatic layer is the Soma layer, and the axon layer is the Axon layer;
[0086] The function and structure of the synaptic layer are as follows: The synaptic layer is mainly responsible for simulating signal transmission between neurons. The original signal is transmitted from a presynaptic neuron to another neuron through the synaptic cleft to achieve effective transmission of information. In this layer, the Sigmoid function is used to simulate the change of the postsynaptic potential. This function is used to determine whether the synapse is excitatory or inhibitory. For the input of the i-th synapse to the j-th synaptic layer, the expression of the Sigmoid function is as follows:
[0087]
[0088] Where i = 1, 2, ..., N; j = 1, 2, ..., M; x i ∈[0,1] represents the i-th input of the synapse; Y ij is the output of the jth synaptic layer; k is a positive constant; weight w ij and threshold θ ij are the required connection parameters;
[0089] The function and structure of the dendritic layer are as follows: The dendritic layer is responsible for processing the output signal from the synaptic layer. This layer processes the combination of multiple input signals by implementing a multiplication function to enhance the model's ability to capture the nonlinear characteristics of the signal. In this layer, the multiplication operator ∏ is used to achieve signal integration. The output of each dendritic branch is formed by the product of all the signals input to the branch, which is specifically expressed as the following formula:
[0090]
[0091] Among them, Z j Represented as the jth dendritic branch;
[0092] The function and structure of the membrane layer are as follows: The membrane layer is mainly responsible for integrating all output signals from the dendritic layer. This layer uses an addition function to process the output signal of each dendritic branch and standardizes the signal through a logical or operation to form a unified output signal. The output V of the membrane layer is obtained by summing the outputs of all dendritic branches. The specific expression is:
[0093]
[0094] Where V represents the output of the membrane layer;
[0095] The function and structure of the somatic layer are as follows: The somatic layer uses the Tanh function to process the output signal from the membrane layer. This function is used to simulate the activation and response process of neurons. When the output signal exceeds the set threshold, the somatic layer generates an activation signal. The output O of the somatic layer is expressed as:
[0096]
[0097] Among them, v represents the membrane layer, s represents the somatic layer, and the parameter k s is a positive number, the threshold θ s Changes from 0 to 1;
[0098] The axon layer will discretely determine the weight of each input of the synaptic layer through the processed somatic layer output.
[0099]
[0100]
[0101]
[0102] Among them, A represents the axon layer, k represents a normal number, std represents the process of calculating the standard deviation, and num represents the number of neurons.
[0103] The connection types and constraints of synapses in the synaptic layer are as follows:
[0104] Case 1: Constant 0 connections, at w ij <0<θ ij Or 0<w ij <θ ij Under the condition, input x i When it changes from 0 to 1, the output is y ij is approximately 0;
[0105] Case 2: Constant 1 connection, at θ ij <0<w ij Or θ ij <w ij <0, input x i When it changes from 0 to 1, the output is y ij About 1;
[0106] Case 3: Excitatory connection, 0<θ ij <w ij Under the condition, input x i When it changes from 0 to 1, the output is Y ij There is a positive correlation;
[0107] Case 4: Inhibitory connection, in w ij <θ ij <0, input x i When changing from 0 to 1, the output is inversely proportional to the input;
[0108] S4: Apply the data obtained from the above processing and analysis to predict chemical disasters. By comparing the actual production data with the model prediction results, adjust the model parameters and prediction strategies to achieve early warning and accurate prediction of possible disaster events in chemical production.
[0109] In order to verify the accuracy of SFDLM for chemical disaster prediction, a large number of experimental analysis and comparison are needed. The present invention compares the hybrid algorithm (ANFIS) based on neural network and fuzzy logic, the long short-term memory network algorithm (LSTM), the multilayer perceptron algorithm (MLP), the support vector computing method (SVM), the neural network algorithm Elman and the seasonal time series prediction model (SARIMA) these 6 algorithms. The data sets used in the experiment of the present invention are 3 real-world chemical production disaster situations, and the data collection in different regions, different influencing factors and different places is selected: Chemical Safety Board Dataset (USA), National Emergency Response Plan Dataset (CHN), Statistics Bureau Dataset (EU), so that the performance of the algorithm is fully trained and tested. For the unification of data, after a series of encapsulation evolutions, the length of the unified data set is 600.
[0110] The performance of SFDLM is affected by three user-defined parameters, including the positive constant k s , threshold θ s As well as the learning rate η in the BP algorithm, the values of each test parameter group are shown in Table 1. In order to obtain reasonable parameter settings in the SFDL model algorithm, the present invention performs three levels of comprehensive experiments on each data set and selects the best parameter combination among all 26 different combinations. In this study, it gives the serial number of the compared parameter groups and models with the minimum prediction error. s ,θ s When selecting η, we make choices according to certain rules. s Start from 1 and end at a certain value based on the prediction result; s The value range is [0, 1], which can be selected in a binary way, and the end value is determined according to the prediction result; the last is the value of η, which is relatively small [0.01, 0.1]. The values given here do not exclude special cases, so it is still necessary to obtain the best results one by one in a binary way. After repeated experiments, the Friedman test is used to determine the best. The Friedman test is a non-parametric statistical test used to identify the test differences of multiple attempts and obtain the best result as k. s =10,θ s =0, η=0.01.
[0111] The present invention uses three indicators to evaluate the ability of each algorithm model, namely, mean square error (MSE), mean absolute percentage error (MAPE) and mean absolute error (MAE). All metrics are calculated between the predicted output and the actual value (de-normalized data).
[0112] Metric 1: Mean Square Error (MSE)
[0113]
[0114] Metric 2: Mean Absolute Percentage Error (MAPE)
[0115]
[0116] Metric 3: Mean Absolute Error (MAE)
[0117]
[0118] Where n is the total number of samples, T i is the actual value, O i is the predicted value.
[0119] The experimental results are shown in Tables 1, 2, and 3:
[0120] Table 1: Comparison of MSE values of seven algorithm models corresponding to the data set
[0121] Dataset SFDLM Elman MLP LSTM ANFIS SVM SARIMA USA 1.50E+02 2.61E+03 3.91E+02 3.62E+02 7.23E+02 1.32E+03 1.37E+04 CHN 2.89E+02 1.02E+03 6.66E+02 5.69E+02 3.82E+02 1.33E+03 1.87E+04 EU 3.15E+02 2.30E+03 9.92E+02 7.47E+02 5.60E+02 1.36E+03 1.76E+04
[0122] Table 2: Comparison of MAPE values of seven algorithm models corresponding to the data set
[0123] Dataset SFDLM Elman MLP LSTM ANFIS SVM SARIMA USA 1.16E-02 1.28E-02 2.30E-02 2.33E-02 3.04E-02 5.02E-02 1.78E-01 CHN 8.98E-03 2.24E-02 2.27E-02 2.57E-02 1.43E-02 4.46E-02 2.07E-01 EU 7.57E-03 2.16E-02 4.16E-02 3.22E-02 2.16E-02 4.43E-02 1.84E-01
[0124] Table 3: Comparison of MAE values of seven algorithm models corresponding to data sets
[0125] Dataset SFDLM Elman MLP LSTM ANFIS SVM SARIMA USA 1.03E+01 3.19E+02 1.64E+01 1.52E+01 1.74E+01 3.36E+01 2.88E+03 CHN 1.56E+01 7.47E+02 2.35E+01 2.01E+01 1.66E+01 3.36E+01 3.53E+03 EU 1.61E+01 7.84E+02 2.74E+01 2.44E+01 1.95E+01 3.39E+01 3.36E+03
[0126] according to Figure 2-Figure 4As shown in the figure, the SFDLM algorithm continues to outperform in all compared indicators (MSE, MAPE, MAE), especially in MSE, with much lower error values than other algorithms. For example, on the US dataset, the MSE of SFDLM is 150, which is much lower than 13700 of SARIMA, showing its high efficiency in data fitting accuracy. On datasets from different regions (such as the United States, China, and the European Union), the SFDLM algorithm maintains a low error rate, indicating that it has good generalization ability and stability. This is particularly important for cross-regional disaster prediction in practical applications. In terms of MAPE and MAE, the SFDLM algorithm also shows low error values, indicating that the accuracy and reliability of its prediction results are higher than other algorithms. The lower MAPE and MAE values indicate its high efficiency in practical applications and can provide more accurate predictions. The SFDLM algorithm can effectively handle various types of datasets and show good performance on multiple different evaluation indicators, which proves its adaptability and flexibility to different data characteristics. Given its excellent performance in multiple indicators, the SFDLM algorithm is not only suitable for chemical disaster prediction, but can also be extended to other fields that require accurate time series prediction, such as financial market analysis and environmental monitoring.
[0127] In summary, the SFDLM algorithm described in the present invention provides a very effective tool for chemical disaster prediction and other fields due to its high accuracy, stability and wide adaptability.
[0128] The above embodiments only express the implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention. It should be pointed out that, for those of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention.
Claims
1. A chemical disaster prediction method based on sequence feature decomposition algorithm, characterized in that: The method comprises the following steps: S1: The SFDL decomposition algorithm is used to process the time series data of chemical production, and the original time series data is decomposed into three main components: seasonal component, trend-cycle component and residual component, where SFDL represents sequence feature decomposition; The SFDL decomposition algorithm for sequence decomposition specifically includes the following steps: S1-1: Decompose the original time series data into three subsequences. The specific calculation formula is: Y k =T k +S k +R k (1) Among them, Y k Indicates the time series data used; S k represents the seasonal term; T k represents the trend term-cycle term; R k represents the residual term; k is 1 to N, representing the length of the sequence; S1-2: SFDL decomposition algorithm uses independent variable x i and the dependent variable y i Regression curve of two parameters Wherein, i=1, 2, ..., n; The fitted value of is the smoothness of y along the scale of the independent variable; S1-3: From x i Select the point r closest to x and assign it a neighborhood weight, and make the following judgments: If r≤n, from {x i } selects the r points closest to x, and then each point will be assigned a local weight according to its distance from x, which is calculated as: Among them, λ r (x) represents the point x farthest from r i distance, w is the cubic weight function, v i (x) is x i The local weight of , t represents the normalized distance; If r>n, λ r The definition of (x) becomes the following formula: Among them, the x closest to x i The weight is the largest; when x i When moving away from x, the weight decreases and eventually drops to 0; finally, when moving away from x i and i Calculate the points of horizontal and vertical coordinates respectively The polynomial function of x is obtained by local fitting value and S2: First, the seasonal component is subjected to Fourier synchronous compression transformation to reduce the noise of the component data. Then, the seasonal component and trend-cycle component obtained by decomposition are further smoothed, and the local weighted regression algorithm is used to optimize and adjust these two components to enhance the adaptive ability to trend and seasonal changes in time series. S3: Combined with the dendritic learning model, a multi-layer neural network model is constructed to perform pattern recognition on the decomposed and smoothed data, including the synaptic layer for signal transmission simulation, the dendritic layer for signal processing, the membrane layer for signal integration and output, and the somatic layer and axon layer for the final output prediction results; The structure of the dendritic learning model is as follows: Synaptic layer: This layer uses the Sigmoid function as a processing tool to simulate the signal transmission between neurons, receives input signals and calculates outputs based on connection weights and thresholds to simulate the excitation or inhibition state of neurons; Dendritic layer: This layer contains multiplication units for receiving input signals from the synaptic layer and processes the combined effects of multiple inputs to enhance the information processing capabilities and nonlinear characteristics of the model; Membrane layer: As the third layer, it is used to integrate the output signals from the dendritic layer and realize signal aggregation through the addition unit to form a response to the complex input of the previous layer; Somatic layer: This layer uses the Tanh function to process the output of the membrane layer and finally generates the output signal of the model, which represents the predicted possibility or risk level of chemical disasters; Axon layer: Determines the weight of each input in the synaptic layer by processing the output of the somatic layer; Among them, the synaptic layer is the Synaptic layer, the dendritic layer is the Dendritic layer, the membrane layer is the Membrane layer, the somatic layer is the Soma layer, and the axon layer is the Axon layer; S4: Apply the data obtained from the above processing and analysis to predict chemical disasters. By comparing the actual production data with the model prediction results, adjust the model parameters and prediction strategies to achieve early warning and accurate prediction of possible disaster events in chemical production.
2. The chemical disaster prediction method based on sequence feature decomposition algorithm according to claim 1 is characterized in that: The specific steps of step S2 are as follows: S2-1: Processing time series data in an iterative detrending loop, which consists of two recursive loops: an inner loop and an outer loop; the inner loop is responsible for updating and smoothing the seasonal component of each generated subsequence; the outer loop is responsible for updating and smoothing the trend component; S2-2: In each iteration, a robustness weight is calculated based on the estimated residual component. This weight is used to reduce the impact of outlier observations in subsequent calculations and improve the model's ability to handle abnormal data. S2-3: Through the above iterative processing, the expression of seasonal and trend components is gradually refined to adapt to the complexity and dynamic changes of time series data in the chemical production process.
3. The chemical disaster prediction method based on sequence feature decomposition algorithm according to claim 1 is characterized in that: The function and structure of the synaptic layer are as follows: the synaptic layer is mainly responsible for simulating signal transmission between neurons. The original signal is transmitted from a presynaptic neuron to another neuron through the synaptic cleft to achieve effective transmission of information. In this layer, the Sigmoid function is used to simulate the change of the postsynaptic potential. This function is used to determine whether the synapse is excitatory or inhibitory. For the i-th synaptic input to the j-th synaptic layer, the expression of the Sigmoid function is as follows: Where i = 1, 2, ..., N; j = 1, 2, ..., M; x i ∈[0,1] represents the i-th input of the synapse; Y ij is the output of the jth synaptic layer; k is a positive constant; weight w ij and threshold θ ij is a required connection parameter.
4. The chemical disaster prediction method based on sequence feature decomposition algorithm according to claim 1 is characterized in that: The function and structure of the dendritic layer are as follows: the dendritic layer is responsible for processing the output signal from the synaptic layer. This layer processes the combination of multiple input signals by implementing a multiplication function to enhance the model's ability to capture the nonlinear characteristics of the signal. In this layer, the multiplication operator ∏ is used to achieve signal integration. The output of each dendritic branch is formed by the product of all the signals input to the branch, which is specifically expressed as the following formula: Among them, Z j Represented as the jth dendritic branch.
5. The chemical disaster prediction method based on sequence feature decomposition algorithm according to claim 1 is characterized in that: The function and structure of the membrane layer are as follows: the membrane layer is mainly responsible for integrating all output signals from the dendritic layer. The layer uses an addition function to process the output signal of each dendritic branch, and standardizes the signal by logical or operation to form a unified output signal; the output V of the membrane layer is obtained by summing the outputs of all dendritic branches, and the specific expression is: Where V represents the output of the membrane layer.
6. The chemical disaster prediction method based on sequence feature decomposition algorithm according to claim 1 is characterized in that: The function and structure of the somatic layer are as follows: the somatic layer uses the Tanh function to process the output signal from the membrane layer, which is used to simulate the activation and response process of neurons. When the output signal exceeds the set threshold, the somatic layer generates an activation signal; the output O of the somatic layer is expressed as: Among them, v represents the membrane layer, s represents the somatic layer, and the parameter k s is a positive number, the threshold θ s Changes from 0 to 1.
7. The chemical disaster prediction method based on sequence feature decomposition algorithm according to claim 1 is characterized in that: The axon layer will discretely determine the weight of each input of the synaptic layer through the processed somatic layer output, that is: Among them, A represents the axon layer, k represents a normal number, std represents the process of calculating the standard deviation, and num represents the number of neurons.
8. The chemical disaster prediction method based on sequence feature decomposition algorithm according to claim 1 is characterized in that: The connection types and constraints of the synapses in the synaptic layer in step S3 are as follows: Case 1: Constant 0 connections, at w ij <0<θ ij or 0 <w ij <θ ij Under the condition, input x i When it changes from 0 to 1, the output is y ij is 0; Case 2: Constant 1 connection, at θ ij <0<w ij Or θ ij <w ij <0, input x i When it changes from 0 to 1, the output is y ij is 1; Case 3: Excitatory connection, when 0 < θ ij <w ij Under the condition, input x i When it changes from 0 to 1, the output is Y ij There is a positive correlation; Case 4: Inhibitory connection, in w ij <θ ij <0, input x i When changing from 0 to 1, the output is inversely proportional to the input.
