Port cargo throughput forecasting method based on evolutionary seasonal trend decomposition

Abstract

The present invention discloses an adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition, aiming to improve the accuracy and stability of port cargo throughput prediction. The time series data is smoothed by local weighted regression technology, and the data is decomposed into seasonal components, trend components and irregular components to capture the key periodic and trend changes in the data, thereby enhancing the adaptability of the model to the intrinsic structure of the data and the accuracy of the prediction. A chaotic particle swarm algorithm based on Gaussian difference variation and multi-scale Levy flight strategy is adopted to optimize the weights and biases of the neural network to improve the stability of the prediction and reduce the prediction error. An adaptive dynamic weight adjustment mechanism is introduced to enable the model to adaptively adjust the prediction strategy according to real-time data, thereby improving the consistency and reliability of long-term predictions. The present invention is suitable for the prediction of port cargo throughput, can effectively support port operation management and planning, and improve decision-making efficiency and accuracy.

Description

Technical Field

[0001] The invention belongs to the field of transportation logistics management and port operation planning, and in particular relates to an adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition. Background Art

[0002] Port cargo throughput refers to the total amount of cargo entering and leaving a port in a certain period of time. It is an important indicator to measure the port's operating efficiency and handling capacity. The accurate prediction of this indicator is extremely important for port planning and daily management, and is directly related to the efficient allocation of resources, the optimization of operational scheduling, and the overall coordination of the supply chain. However, port cargo throughput is affected by many factors such as global trade liquidity, economic activity levels, policy changes, and unexpected events, and may fluctuate significantly. Therefore, there are many challenges in accurately predicting port cargo throughput.

[0003] At present, a variety of methods have been used to predict port cargo throughput, mainly including: prediction models based on neural networks: Some researchers have developed models based on back-propagation neural networks, using economic indicators such as total foreign trade imports and exports, industrial output value, etc. as inputs to try to capture the dynamic changes in cargo throughput. Time series analysis methods: Other researchers focus on using time series analysis techniques, such as Elman neural networks, which can process dynamic information and try to identify patterns and trends from historical data.

[0004] Although these methods perform well in some cases, they usually have the following shortcomings: Insufficient complex data processing capabilities: Existing methods often have difficulty in effectively processing complex data containing a lot of noise and nonlinear characteristics, especially in scenarios with large data volumes and rapidly changing environments. Insufficient consideration of long-term trends and cyclical factors: Many models fail to fully consider the long-term trends and cyclical changes in port cargo throughput data, which limits the accuracy and reliability of predictions. Insufficient handling of uncertainties: Port cargo throughput is affected by a variety of external factors, including economic fluctuations, policy changes and other emergencies. Existing technologies have failed to effectively deal with these uncertainties, resulting in deviations in prediction results. Summary of the invention

[0005] The purpose of the present invention is to address the defects of the prior art and provide an adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition, so as to reduce the error of port cargo throughput prediction, thereby improving the prediction accuracy and overall planning efficiency, and solving the problems of port operation efficiency and port cargo throughput prediction.

[0006] To achieve the above object, the present invention specifically provides the following technical solution: an adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition, which uses the following steps to complete the prediction of port cargo throughput:

[0007] S1: Data collection and processing: Collect historical port cargo throughput data, which specifically includes time series data and trade flows; clean, standardize and preprocess the data. The preprocessing specifically uses normalization processing, and the calculation formula for normalization processing is:

[0008]

[0009] Among them, y(x i ) represents the time point x i The original data value, y min and max Represent the minimum and maximum values ​​in the data set, y(x i ) normalized is the result of normalization;

[0010] S2: Model construction: Using the evolutionary seasonal trend decomposition algorithm, the port cargo throughput Y x Decomposed into seasonal terms S x 、Trend item T x and the residual term I x ;

[0011] S3: Algorithm optimization: Use the improved particle swarm algorithm to optimize the weights and biases of the feedforward neural network, conduct in-depth analysis of the residual terms, and continuously iterate to find the optimal solution through mutation and optimization strategies, thereby reducing the prediction error and improving the stability of the prediction;

[0012] S4: Dynamic weight adjustment: Introduce a dynamic weight adjustment mechanism to adaptively adjust the weights of each component based on data and optimize the final prediction model.

[0013] S5: Result evaluation and verification: The prediction results of the model are evaluated using mean square error, mean absolute error and mean absolute percentage error evaluation indicators.

[0014] As a preferred technical solution of the present invention, the evolutionary seasonal trend decomposition algorithm in step S2 is specifically a seasonal trend decomposition algorithm based on LOESS, and its algorithm flow specifically includes the following steps, wherein LOESS stands for local weighted regression:

[0015] S2-1: Original sequence decomposition: Use the LOESS algorithm to decompose the original time series Y x Decompose it, the specific calculation formula is:

[0016] Yx = S x + T x + I x , x = 1, 2, 3, …, n (3)

[0017] Among them, S x represents the seasonal component, T x represents the trend component, I x represents the irregular component;

[0018] S2-2: Local regression fitting: When calculating the local fitting curve f(x), with the signal point x as the center, select the set of p points closest to x and assign a local weight v i (x) to them according to the distance between each point and x. Its calculation formula is:

[0019]

[0020]

[0021]

[0022] Among them, p < n, W is a cubic weight function for linearly weighting adjacent data sets λ p (x) represents the distance scaling parameter, which is used to adjust how the distance influence between points is converted into weights, scaling the distance from x to the farthest point x i ;

[0023] S2-3: Iterative update and robust weights: The decomposition process uses LOESS to smooth the sequence through an inner loop and an outer loop, decomposing the seasonal component and the trend component, and continuously updating and iterating in the inner loop. During the iteration process, robust weights ε t are formed according to the estimated residual component. Its calculation formula is as follows:

[0024]

[0025]

[0026] σ = 6median(|I x |) (9)

[0027] Among them, B(b) is a threshold function for calculating the robust weight ε x ; σ is the estimated value of the standard deviation, and 6median represents the clustering algorithm. By adjusting the robust weight, the interference of outliers to the result can be effectively reduced.

[0028] As a preferred technical solution of the present invention, the algorithm optimization described in step S3 adopts a chaotic particle swarm algorithm improved by Gaussian difference mutation and multi-scale Levy flight strategy to optimize the feedforward neural network, and the specific steps of the algorithm optimization include: S3-1: using Logistics mapping to initialize the population distribution; S3-2: proposing a multi-scale Levy flight strategy to update the particle position; S3-3: using Gaussian difference mutation strategy to perform probabilistic mutation on the particle position.

[0029] As a preferred technical solution of the present invention, the mapping formula for initializing the population distribution of the Logistics mapping described in step S3-1 is:

[0030] X i+1 =chaos*X i *(1-x i ) (10)

[0031] Among them, X i is the Logistics mapping value in the current iteration step; X i+1 is the Logistics mapping value of the next iteration step; chaos is the chaos coefficient. When chaos takes different values, the Logistics mapping has rich dynamic behaviors. The formula for initializing the population using the Logistics mapping is:

[0032] POS i,j =lb j +(ubx-lb j )×Logistics i,j (11)

[0033] Among them, pos i,j is the initial position of the ith particle in the jth dimension, lb j andub j are the lower and upper bound vectors of the search space respectively.

[0034] As a preferred technical solution of the present invention, the calculation process of updating the particle position by the multi-scale Levy flight strategy described in step S3-2 is:

[0035] Generate a random number δ in the interval [0,1] that follows a uniform distribution. If δ<0.5, the particle position update formula is:

[0036]

[0037]

[0038] in, represents the velocity vector of particle i at the next time step t+1, ω is the inertia weight, c1 and c 2 is the acceleration factor, r 1 and r 2 is a random number in the range [0,1], pbest i is the optimal value of particle i in the search process, is the position of particle i at time t, gbest is the optimal value of all particles in the search process; if δ>0.5, the multi-scale Levy flight strategy is used to update the particle position, which is specifically simulated by the Mantegna algorithm. The mathematical expression of the Mantegna algorithm is:

[0039]

[0040]

[0041]

[0042]

[0043]

[0044] Where s is the step size, μ i and v i is a normal distribution, β represents the exponential parameter of the stable distribution, which is taken as a constant of 1.5. The particle position is updated using the multi-scale Levy flight strategy, Γ is the gamma function, and the update calculation formula is:

[0045]

[0046]

[0047] Where λ is the multi-scale scaling factor, f i is the step size control factor, T is the total number of iterations, and t is the current number of iterations.

[0048] As a preferred technical solution of the present invention, the Gaussian distribution function formula for probabilistically mutating the particle position using the Gaussian difference mutation strategy described in step S3-3 is:

[0049]

[0050] Among them, x is the value of the random variable, μ is the mean of the distribution, and σ is the standard deviation of the distribution. Gaussian difference mutation first calculates the difference vector between the current individual and the global optimal solution gbest, and then adds the random perturbation term of the Gaussian distribution to generate the mutated particle position Mutantpos. The specific update calculation formula is:

[0051] Mutantpos=pos(i,:)+gaussian·(gbest-pos(i,:)) (22)

[0052] Among them, gaussian represents the random perturbation term, Mutantpos represents the position of the mutated particle, and pos(i,:) represents the position vector of the i-th particle.

[0053] As a preferred technical solution of the present invention, the adaptive dynamic weight adjustment in step S4 specifically adopts a weighted fusion module, which completes the weighted fusion of the prediction results through the following steps:

[0054] S4-1: Seasonal component S x Moving average is used for further processing to reduce random noise. The specific calculation formula is:

[0055] S x =AvgPool(Padding(S)) (23)

[0056] Among them, AvgPool represents average pooling, and Padding represents padding operation;

[0057] S4-2: Trend component T x The appropriate trend curve is obtained by nonlinear fitting of cubic polynomials. The specific calculation formula is:

[0058] T x =at 3 +bt 2 +ct+d (24)

[0059] Among them, a, b, c, d represent the coefficients of the polynomial, and t represents the input data;

[0060] S4-3: Residual component I x GLPSO-FNN is used for prediction, where GLPSO-FNN stands for feedforward neural network optimized by chaotic particle swarm algorithm improved by Gaussian difference mutation and Levy flight strategy;

[0061] S4-4: Adaptively weight the three component subsequences to form the final port cargo throughput prediction result. The specific calculation formula is:

[0062] Y x =W S S x +W T T x +W I I x (25)

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071] Among them, Y x represents the throughput prediction result, represents the actual value, f is the loss function, W S , W T and W I represents the weight parameter, initialized to 1, N is the number of samples, τ is the error feedback adjustment coefficient, e t-1 is the error of the previous step, E is the error set, which is used to normalize the error, η is the learning rate, which is used to control the step size of the weight update and is set to 1e-2, μ is the smoothing factor, ΔW pre It is the weight update of the previous step, which dynamically adjusts the weights adaptively through error feedback to improve the overall prediction accuracy and adaptability of the model.

[0072] Compared with the prior art, the present invention has the following beneficial effects:

[0073] The evolutionary seasonal trend decomposition algorithm (ESTD) adopted by the present invention significantly improves the accuracy of the prediction model. By carefully decomposing the port cargo throughput data into seasonal components, trend components and irregular components, ESTD can more deeply analyze and understand the inherent structure of the data. This decomposition helps the model to accurately capture the cyclical changes and long-term trends that affect the throughput, thereby reducing the prediction error and providing a more accurate estimate of future throughput. Comparative experiments and long-term tracking show that the ESTD model shows excellent stability and consistency in prediction. Over time, despite the fluctuations in data and changes in the external environment, the model can continue to provide prediction results that are highly consistent with the real data sequence. This stability is particularly suitable for data such as port cargo throughput that is affected by multiple factors and susceptible to external disturbances, ensuring the reliability of long-term planning and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1A flowchart of the adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition provided by the present invention;

[0075] Figure 2 A visualization forecast result diagram of cargo throughput of coastal ports above designated size according to an embodiment of the present invention;

[0076] Figure 3 A visualization forecast result diagram of foreign trade cargo throughput of coastal ports above designated size according to an embodiment of the present invention;

[0077] Figure 4 A visualization forecast result diagram of domestic trade cargo throughput in coastal ports above a certain scale is provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0078] 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.

[0079] Embodiment: The adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition disclosed in the present invention has the following specific steps:

[0080] S1: Data collection and processing: Collect historical port cargo throughput data, which specifically includes time series data and trade flows; clean, standardize and preprocess the data. The preprocessing specifically uses normalization processing, and the calculation formula for normalization processing is:

[0081]

[0082] Among them, y(x i ) represents the time point x i The original data value, y min and max Represent the minimum and maximum values ​​in the data set, y(x i ) normalized is the result of normalization;

[0083] S2: Model construction: Using the Evolutionary Seasonal Trend Decomposition (ESTD) algorithm, the port cargo throughput Y x Decomposed into seasonal terms S x 、Trend item T x and the residual term I x ;

[0084] The evolutionary seasonal trend decomposition algorithm is specifically a seasonal trend decomposition algorithm based on LOESS, and its algorithm flow specifically includes the following steps, where LOESS stands for locally weighted regression:

[0085] S2-1: Original sequence decomposition: Use the LOESS algorithm to decompose the original time series Y x and the specific calculation formula is:

[0086] Y x = S x + T x + I x , x = 1, 2, 3, …, n (3)

[0087] where S x represents the seasonal component, T x represents the trend component, and I x represents the irregular component;

[0088] S2-2: Local regression fitting: When calculating the local fitting curve f(x), with the signal point x as the center, select the set of p points closest to x and assign a local weight v i (x) to them according to the distance between each point and x, and its calculation formula is:

[0089]

[0090]

[0091]

[0092] where p < n, W is a cubic weight function to linearly weight the adjacent data sets and λ p (x) represents the distance scaling parameter, which is used to adjust how the distance influence between points is converted into weights, scaling the distance from x to the farthest point x i ;

[0093] S2-3: Iterative update and robust weights: The decomposition process uses LOESS to smooth the sequence through an inner loop and an outer loop, decomposing the seasonal component and the trend component, and continuously updating and iterating in the inner loop. During the iteration process, robust weights ε t are formed according to the estimated residual component, and its calculation formula is as follows:

[0094]

[0095]

[0096] σ = 6median(|I x |) (9)

[0097] where B(b) is a threshold function used to calculate the robust weight εx ; σ is the estimated value of the standard deviation, 6median represents the clustering algorithm, and by adjusting the robustness weight, the interference of outliers on the results can be effectively reduced.

[0098] S3: Algorithm optimization: Use the improved particle swarm algorithm to optimize the weights and biases of the feedforward neural network, conduct in-depth analysis of the residual terms, and continuously iterate to find the optimal solution through mutation and optimization strategies, thereby reducing the prediction error and improving the stability of the prediction;

[0099] The algorithm optimization uses Gaussian difference mutation and multi-scale Levy flight strategy to improve the chaotic particle swarm algorithm to optimize the feedforward neural network. The specific steps of the algorithm optimization include:

[0100] S3-1: Initialize population distribution using Logistics mapping;

[0101] The mapping formula for initializing the population distribution of Logistics mapping is:

[0102] X i+1 =chaos*X i *(1-X i ) (10)

[0103] Among them, X i is the Logistics mapping value in the current iteration step; X i+1 is the Logistics mapping value of the next iteration step; chaos is the chaos coefficient. When chaos takes different values, the Logistics mapping has rich dynamic behaviors. The formula for initializing the population using the Logistics mapping is:

[0104] POS i,j =lb j +(ub j -lb j )×Logistics i,j (11)

[0105] Among them, pos i,j is the initial position of the ith particle in the jth dimension, lb j andub j are the lower and upper bound vectors of the search space respectively;

[0106] S3-2: Propose a multi-scale Levy flight strategy to update the particle position;

[0107] The calculation process of the multi-scale Levy flight strategy to update the particle position is proposed as follows:

[0108] Generate a random number δ in the interval [0,1] that follows a uniform distribution. If δ<0.5, the particle position update formula is:

[0109]

[0110]

[0111] in, represents the velocity vector of particle i at the next time step t+1, ω is the inertia weight, c 1 and c 2 is the acceleration factor, r 1 and r 2 is a random number in the range [0,1], pbest i is the optimal value of particle i in the search process, is the position of particle i at time t, gbest is the optimal value of all particles in the search process; if δ>0.5, the multi-scale Levy flight strategy is used to update the particle position, which is specifically simulated by the Mantegna algorithm. The mathematical expression of the Mantegna algorithm is:

[0112]

[0113]

[0114]

[0115]

[0116]

[0117] Where s is the step size, μ i and v i is a normal distribution, β represents the exponential parameter of the stable distribution, which is taken as a constant of 1.5. The particle position is updated using the multi-scale Levy flight strategy, Γ is the gamma function, and the update calculation formula is:

[0118]

[0119]

[0120] Where λ is the multi-scale scaling factor, f i is the step size control factor, T is the total number of iterations, and t is the current number of iterations;

[0121] S3-3: Use Gaussian difference mutation strategy to perform probabilistic mutation on particle positions;

[0122] The Gaussian distribution function formula for probabilistically mutating the particle position using the Gaussian difference mutation strategy is:

[0123]

[0124] Among them, x is the value of the random variable, μ is the mean of the distribution, and σ is the standard deviation of the distribution. Gaussian difference mutation first calculates the difference vector between the current individual and the global optimal solution gbest, and then adds the random perturbation term of the Gaussian distribution to generate the mutated particle position Mutantpos. The specific update calculation formula is:

[0125] Mutantpos=pos(i,:)+gaussian·(gbest-pos(i,:)) (22)

[0126] Among them, gaussian represents the random perturbation term, Mutantpos represents the position of the mutated particle, and p os(i,:) represents the position vector of the i-th particle;

[0127] S4: Dynamic weight adjustment: Introduce a dynamic weight adjustment mechanism to adaptively adjust the weights of each component based on data and optimize the final prediction model.

[0128] The dynamic weight adjustment specifically adopts a weighted fusion module, which completes the weighted fusion of the prediction results through the following steps:

[0129] S4-1: Seasonal component S x Moving average is used for further processing to reduce random noise. The specific calculation formula is:

[0130] S x =AvgPool(Padding(S)) (23)

[0131] Among them, AvgPool represents average pooling, and Padding represents padding operation;

[0132] S4-2: Trend component T x The appropriate trend curve is obtained by nonlinear fitting of cubic polynomials. The specific calculation formula is:

[0133] T x =at 3 +bt 2 +ct+d (24)

[0134] Among them, a, b, c, d represent the coefficients of the polynomial, and t represents the input data;

[0135] S4-3: Residual component I xGLPSO-FNN is used for prediction, where GLPSO-FNN stands for feedforward neural network optimized by chaotic particle swarm algorithm improved by Gaussian difference mutation and multi-scale Levy flight strategy;

[0136] S4-4: Adaptively weight the three component subsequences to form the final port cargo throughput prediction result. The specific calculation formula is:

[0137] Y x =W S S x +W T T x +W I I x (25)

[0138]

[0139]

[0140]

[0141]

[0142]

[0143]

[0144]

[0145]

[0146] Among them, Y x represents the throughput prediction result, represents the actual value, f is the loss function, W S , W T and W I represents the weight parameter, initialized to 1, N is the number of samples, τ is the error feedback adjustment coefficient, e t-1 is the error of the previous step, E is the error set, which is used to normalize the error, η is the learning rate, which is used to control the step size of the weight update and is set to 1e-2, μ is the smoothing factor, ΔW pre It is the weight update of the previous step, which dynamically adjusts the weights adaptively through error feedback to improve the overall prediction accuracy and adaptability of the model;

[0147] S5: Result evaluation and verification: Use evaluation indicators such as mean square error (MSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) to evaluate the prediction results of the model. Through comparative analysis, verify the prediction ability and practicality of the model.

[0148] In order to verify the accuracy of the ESTD algorithm in predicting port cargo throughput, a large number of numerical experimental analysis and comparison are required. The present invention compares the five models of Elman neural network, multi-layer perceptron (MLP), long short-term memory (LSTM), support vector machine (SVM), and adaptive neural fuzzy system (ANFIS) in the literature. Since there is currently no standard test set for the port cargo throughput prediction problem, the present invention specially selects the monthly data (2005 to 2023) of the cargo throughput of coastal ports above a certain scale, foreign trade cargo throughput, and domestic trade cargo throughput released by the Ministry of Transport of China, and uses 70% of the data for model training and 30% of the data for testing.

[0149] The present invention uses different models to obtain the performance of three key evaluation indicators, namely, mean square error (MSE), mean absolute error (MAE) and mean absolute percentage error (MAPE), and the correlation coefficient (R) on three data sets.

[0150] Metric 1: MSE:

[0151] Metric 2: MAE:

[0152] Metric 3: MAPE:

[0153] Indicator 4: R:

[0154] Where n is the total number of samples, T i is the actual value, O i is the predicted value.

[0155] The experimental results are shown in Table 1:

[0156] Table 1: Experimental results of 6 models on each dataset

[0157]

[0158] As can be seen from Table 1, under the comparative experimental results of different models, the performance of the algorithm proposed in the present invention in terms of MSE, MAE, MAPE and R indicators is better than that of the comparison model, showing lower prediction error and stronger positive correlation, and has good stability and consistency. Figure 2-4It can be seen that as time goes by, the throughput shows an overall upward trend. ESTD accurately predicts the long-term trend of the data, which is infinitely close to the original sequence. However, other comparison algorithms have large errors and are greatly lacking in accuracy. And as time goes by, the stability of the prediction is getting lower and lower. The ESTD model proposed in the present invention can more effectively reduce the error of port cargo throughput prediction, thereby improving the prediction accuracy and overall planning efficiency, and solving the problems of port operation efficiency and port cargo throughput prediction.

[0159] Through the above embodiments, the prediction scheme proposed by the present invention can predict the future port cargo throughput with high accuracy, and provide a scientific basis for port management and operation decisions.

[0160] 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. An adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition, characterized in that: The method includes the following steps: S1: Data collection and processing: Collect historical port cargo throughput data, which specifically includes time series data and trade flows; clean, standardize and preprocess the data. The preprocessing specifically uses normalization processing, and the calculation formula for normalization processing is: Among them, y(x i ) represents the time point x i The original data value, y min and max Represent the minimum and maximum values ​​in the data set, y(x i ) normalized is the result of normalization; S2: Model construction: Using the evolutionary seasonal trend decomposition algorithm, the port cargo throughput Y x Decomposed into seasonal terms S x 、Trend item T x and the residual term I x ; S3: Algorithm optimization: Use the improved particle swarm algorithm to optimize the weights and biases of the feedforward neural network, conduct in-depth analysis of the residual terms, and continuously iterate to find the optimal solution through mutation and optimization strategies, thereby reducing the prediction error and improving the stability of the prediction; S4: Dynamic weight adjustment: Introduce a dynamic weight adjustment mechanism to adaptively adjust the weights of each component based on data-driven, and optimize the final prediction model; Dynamic weight adjustment specifically adopts a weighted fusion module, which completes the weighted merging of prediction results through the following steps: S4-1: Seasonal component S x Moving average is used for further processing to reduce random noise. The specific calculation formula is: S x =AvgPool(Padding(S)) (23) Among them, AvgPool represents average pooling, and Padding represents padding operation; S4-2: Trend component T x The appropriate trend curve is obtained by nonlinear fitting of cubic polynomials. The specific calculation formula is: T x =al 3 +bl 2 +cl+d (24) Among them, a, b, c, d represent the coefficients of the polynomial, and l represents the input data; S4-3: Residual component I x GLPSO-FNN is used for prediction, where GLPSO-FNN stands for feedforward neural network optimized by chaotic particle swarm algorithm improved by Gaussian difference mutation and Levy flight strategy; S4-4: Adaptively weight the three component subsequences to form the final port cargo throughput prediction result. The specific implementation formula is as follows: Y x =W s S x +W T T x +W I I x (25) Among them, Y x represents the throughput prediction result, represents the actual value, f is the loss function, W S , W T and W I represents the weight parameter, initialized to 1, N is the number of samples, τ is the error feedback adjustment coefficient, e t-1 is the error of the previous step, E is the error set, which is used to normalize the error, η is the learning rate, which is used to control the step size of weight update, set to 1e-2, ψ is the smoothing factor, ΔW pre It is the weight update of the previous step, which dynamically adjusts the weights adaptively through error feedback to improve the overall prediction accuracy and adaptability of the model; S5: Result evaluation and verification: The prediction results of the model are evaluated using mean square error, mean absolute error and mean absolute percentage error evaluation indicators.

2. The adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition according to claim 1 is characterized in that: The evolutionary seasonal trend decomposition algorithm in step S2 is specifically a seasonal trend decomposition algorithm based on LOESS, and its algorithm flow specifically includes the following steps, where LOESS stands for local weighted regression: S2-1: Original sequence decomposition: Use the LOESS algorithm to decompose the original time series Yx. The specific calculation formula is: Y x =S x +T x +I x ,x=1,2,3,...,n (3) Among them, S x represents the seasonal component, T x represents the trend component, I x represents the residual component; S2-2: Local regression fitting: When calculating the local fitting curve f(x), take the signal point x as the center and select the set of p points closest to x And assign a local weight v to each point according to its distance from x i (x), which is calculated as: Where p < n, W is a cubic weight function for adjacent data sets Linear weighting, λ p (x) represents the distance scaling parameter, which is used to adjust how the distance between points is converted into weights, scaling x to the farthest point x i distance; S2-3: Iterative update and robust weight: The decomposition process uses LOESS to smooth the sequence through inner and outer loops, decompose the seasonal component and trend component, and continuously update and iterate in the inner loop. In the iterative process, the robust weight ε is formed according to the estimated residual component x , and its calculation formula is as follows: σ=6median(|I x |) (9) Among them, B(b) is a threshold function used to calculate the robustness weight ε x ; σ is the estimated value of the standard deviation, 6median represents the clustering algorithm, and by adjusting the robustness weight, the interference of outliers on the results can be effectively reduced.

3. The adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition according to claim 1 is characterized in that: The algorithm optimization described in step S3 uses a chaotic particle swarm algorithm improved by Gaussian difference mutation and multi-scale Levy flight strategy to optimize the feedforward neural network. The specific steps of the algorithm optimization include: S3-1: using Logistics mapping to initialize the population distribution; S3-2: proposing a multi-scale Levy flight strategy to update the particle position; S3-3: using a Gaussian difference mutation strategy to perform probabilistic mutation on the particle position.

4. The adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition according to claim 3 is characterized in that: The mapping formula for the initialization population distribution of the Logistics mapping described in step S3-1 is: X i+1 =chaos*X i *(1-X i ) (10) Among them, X i is the Logistics mapping value in the current iteration step; X i+1 is the Logistics mapping value of the next iteration step; chaos is the chaos coefficient. When chaos takes different values, the Logistics mapping has rich dynamic behaviors. The formula for initializing the population using the Logistics mapping is: pos i，j =lb j +(ub j -lb j )×Logistics i，j (11) Among them, pos i，j is the initial position of the ith particle in the jth dimension, lb j andub j are the lower and upper bound vectors of the search space respectively.

5. The adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition according to claim 3 is characterized in that: The calculation process of the proposed multi-scale Levy flight algorithm to update the particle position described in step S3-2 is: Generate a random number δ in the interval [0, 1] that follows a uniform distribution. If δ<0.5, the particle position update formula is: in, represents the velocity vector of particle i at the next time step t+1, ω is the inertia weight, c1 and c2 are acceleration factors, r1 and r2 are random numbers in the range [0, 1], and pbest i is the optimal value of particle i in the search process, is the position of particle i at time t, gbest is the optimal value of all particles in the search process; if δ>0.5, the multi-scale Levy flight strategy is used to update the particle position, which is specifically simulated by the Mantegna algorithm. The mathematical expression of the Mantegna algorithm is: Where s is the step size, μi and ν i is a normal distribution, β represents the exponential parameter of the stable distribution, which is taken as a constant of 1.

5. The particle position is updated using the multi-scale Levy flight strategy, Γ is the gamma function, and the update calculation formula is: Among them, λ is the multi-scale scaling factor, T is the total number of iterations, z is the current iteration number, and fi is the step size control factor.

6. The adaptive port cargo throughput prediction method based on evolutionary seasonal trend decomposition according to claim 3 is characterized in that: The Gaussian distribution function formula for probabilistically mutating the particle position using the Gaussian difference mutation strategy described in step S3-3 is: Among them, x is the value of the random variable, μ is the mean of the distribution, and σ is the standard deviation of the distribution. Gaussian difference mutation first calculates the difference vector between the current individual and the optimal value gbest of all particles in the search process, and then adds the random perturbation term of Gaussian distribution to generate the mutated particle position Mutantpos. The specific update calculation formula is: Mutantpos=pos(i,:)+gaussian·(gbest-pos(i,:)) (22) Among them, gaussian represents the random perturbation term, Mutantpos represents the position of the mutated particle, and pos(i,:) represents the position vector of the i-th particle.

Images