Intelligent warehouse location planning optimization method based on hybrid particle swarm optimization

Abstract

The invention discloses an intelligent warehouse location planning optimization method based on a hybrid particle swarm algorithm, and the method comprises the steps: 1, building a warehouse space model based on a Manhattan distance, and designing an overall fitness function; step 2, establishing a warehouse-in and warehouse-out frequency prediction model based on the long short-term memory network; 3, performing random disturbance and stretching processing on a prediction result of the warehouse in-out frequency prediction model, and expanding a frequency distribution range; step 4, based on a particle swarm algorithm, through self-adaptive adjustment of inertia weight, realizing parking space optimization dynamic search and convergence balance; 5, based on a genetic algorithm, introducing crossover and mutation operation in a particle updating process, and enhancing the diversity of storage location distribution; and 6, based on a simulated annealing algorithm, by introducing temperature parameters and an inferior solution acceptance mechanism, jumping out of a local optimal solution, and improving the global storage location optimization capability. According to the invention, the length and time consumption of a material circulation path are effectively reduced, and the operation efficiency and management capability of the intelligent warehousing system are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent warehouse location planning, and relates to an intelligent warehouse location planning optimization method based on a hybrid particle swarm optimization algorithm. Background Art

[0002] The automated production process based on warehouse logistics planning is crucial for the development of various industries. In this context, the optimization of intelligent warehouse location planning has received extensive attention. Its key points lie in the establishment of a warehouse space model, the prediction of inbound and outbound frequencies, and intelligent location planning. The purpose of establishing a warehouse space model is to establish a quantitative standard for location planning, the prediction of inbound and outbound frequencies is to predict future inbound and outbound data, and the purpose of intelligent location planning is to reasonably allocate locations and improve warehouse efficiency.

[0003] In the research process of intelligent warehouse location planning, the following problems are found: (1) Existing methods often concentrate goods with high inbound and outbound frequencies in locations closer to the entrance and exit, which easily leads to the accumulation of high-frequency goods in local areas, resulting in uneven utilization of warehouse space, and prone to congestion and path conflicts; (2) Traditional location planning models are mostly optimized once based on static data and lack the ability to predict future demand changes; (3) It is easy to fall into local optimal solutions, especially in high-dimensional spaces or under complex constraint conditions, and it is difficult to guarantee the search efficiency and the quality of solutions. Summary of the Invention

[0004] In order to solve the above technical problems existing in the prior art, the present invention proposes an intelligent warehouse location planning optimization method based on a hybrid particle swarm optimization algorithm, and its specific technical solution is as follows:

[0005] An intelligent warehouse location planning optimization method based on a hybrid particle swarm optimization algorithm, comprising:

[0006] Step 1: Establish a warehouse space model based on Manhattan distance and design an overall fitness function;

[0007] Step 2: Establish an inbound and outbound frequency prediction model based on an LSTM long short-term memory network;

[0008] Step 3: Perform random perturbation and stretching processing on the prediction results of the inbound and outbound frequency prediction model to expand the frequency distribution range;

[0009] Step 4: Based on the particle swarm optimization algorithm, achieve dynamic search and convergence balance of location optimization by adaptively adjusting the inertia weight;

[0010] Step 5: Based on the genetic algorithm, introduce crossover and mutation operations during particle update to enhance the diversity of location distribution;

[0011] Step 6: Based on the simulated annealing algorithm, by introducing a temperature parameter and a mechanism for accepting inferior solutions, jump out of the local optimal solution and improve the global storage location optimization ability.

[0012] Further, in Step 1, a three-dimensional coordinate grid model of the storage space is established, the Manhattan distances from the outbound and inbound positions to each storage location are calculated, and a height penalty term is introduced to quantify the transfer cost of materials in the warehouse.

[0013] The formula for calculating the Manhattan distance is:

[0014] D = |x - x0| + |y - y0| + |h - h0|,

[0015] where (x0, y0, h0) is the position of the outbound or inbound port, (x, y, h) is the three-dimensional coordinate of the target storage location, and D is the Manhattan distance between the entrance / exit and the target storage location;

[0016] When calculating the Manhattan distance in the warehouse space model, a height penalty term is introduced:

[0017] D′ = D + β·h,

[0018] where β is the height penalty coefficient and h is the height of the storage location;

[0019] A frequency dispersion penalty term is also introduced, and the calculation formula is:

[0020]

[0021] where λ is the penalty coefficient used to control the influence weight of the penalty term in the overall fitness function, N is the number of stock keeping units (SKUs), α is the sensitivity coefficient of the frequency difference, which controls the convergence speed of the penalty term, |f i -f j | is the outbound frequency difference between SKUi and SKUj, and D(i, j) is the Manhattan distance between SKUi and SKUj; when the outbound and inbound frequencies of SKUi and SKUj are close, the value of the penalty term is large, prompting the algorithm to store them separately; when the outbound and inbound frequencies of SKUi and SKUj are quite different, the penalty term tends to 0, and the SKUs with large frequency differences are relatively concentrated. When the storage location distance D(i, j) increases, the penalty term decreases, encouraging SKUs with similar frequencies to be distributed farther apart in space; the penalty term increases rapidly as the frequency difference decreases and the distance decreases.

[0022] Further, in Step 1, the overall fitness function is:

[0023]

[0024] where f out,i and f in,iis the inbound and outbound frequency of SKUi, D out,i and D in,i is the Manhattan distance from the SKU to the entrance and exit, w i and v i are the weight and volume of the SKU, and h is the height of the SKU in the storage space.

[0025] Furthermore, the specific steps of step 2 include:

[0026] First, in the data preprocessing stage, using multi-dimensional historical frequency data, normalize the historical frequency data, and the normalization formula is:

[0027]

[0028] Among them, f' i,j and g' i,j are the normalized inbound and outbound frequencies respectively, f i,j and g i,j are the original frequencies of the i-th time step and the j-th SKU respectively, min(f j ) and max(f j ) are the minimum and maximum outbound frequencies of the j-th stock keeping unit (SKU) among all time steps; after normalization, all frequency data are compressed into the interval [0, 1];

[0029] Next, divide the historical frequency data into time windows to construct a sliding time series dataset:

[0030]

[0031] Among them, X (k) is the input sequence of the k-th time window, with a dimension of τ×N, and y (k) is the predicted frequency at the subsequent moment of the corresponding time window;

[0032] Adopt a multi-layer LSTM structure to establish an inbound and outbound prediction model, where the hidden layer is used to extract the time features of frequency changes, and the output layer uses a linear activation layer to output the predicted values of the inbound and outbound frequencies of the SKU at future moments. The predicted future frequency distribution is:

[0033]

[0034] Among them, W out is the weight of the last layer, b out is the bias, and h t is the hidden state of the current time step;

[0035] Optimize the model parameters through adaptive learning rate adjustment and mean square error loss function. The formula is:

[0036]

[0037] Among them, is the predicted frequency, f i is the true frequency;

[0038] The optimization of the inbound and outbound prediction model uses the Adam adaptive learning rate algorithm to adjust the parameters.

[0039] Furthermore, the specific steps of step 3 include:

[0040] First, introduce random perturbations by superimposing Gaussian noise on the prediction results to simulate the possible volatility and uncertainty of future frequencies, thus avoiding overly regular frequency distributions;

[0041] Subsequently, introduce a stretching coefficient to expand the frequency distribution range, moderately amplify or reduce some frequency values to increase the difference and diversity of the frequency distribution;

[0042] After combining the random perturbations and stretching processing, the future frequency distribution is:

[0043]

[0044] Among them, 1+∈ is the random fluctuation coefficient, and S is the stretching coefficient.

[0045] Furthermore, in step 4, the particle swarm algorithm continuously adjusts its position and velocity by simulating the search behavior of particles in the solution space to seek the optimal storage location distribution. The velocity update formula and position update formula of the particles are respectively:

[0046] v i (t + 1) = w(t)·v i (t) + c1·r1·(p best,i -x i (t)) + c2·r2·(g best -x i (t)),

[0047] x i (t + 1) = x i (t) + v i (t + 1),

[0048] Among them, x i (t) is the storage location distribution of the i-th particle at the t-th iteration, v i (t) is the velocity of the i-th particle, p best,i is the optimal solution in the history of the i-th particle, g bestis the global optimal solution in the population history, w(t) is the inertia weight, c1 and c2 are learning factors representing acceleration coefficients, and r1, r2 are random numbers in the range of [0, 1][0, 1][0, 1].

[0049] Furthermore, the velocity update of the particle adopts a probability mapping mechanism, mapping the velocity of the particle to a probability through the Sigmoid function:

[0050]

[0051] Among them, P i,j is the jump probability of particle i in dimension j, and v i,j (t) is the velocity of the i-th particle in the j-th dimension.

[0052] Furthermore, in step 4, an adaptive inertia weight mechanism is introduced, and the inertia weight formula is:

[0053]

[0054] Among them, w max is the initial inertia weight, set to 0.90 to enhance the exploration ability, and w min is the final inertia weight, set to 0.40 to enhance the convergence ability; t is the current iteration number, and T is the maximum iteration number; in the initial stage of iteration, a larger inertia weight is adopted to enhance the global exploration ability, so as to conduct a wider search in the storage location space and prevent falling into local optima; in the later stage of iteration, the inertia weight is gradually reduced to accelerate the convergence speed of the algorithm, so as to finely adjust the storage location distribution and ensure the optimization of the access efficiency of the stock keeping unit (SKU).

[0055] Furthermore, in step 5, a genetic algorithm (GA) is integrated on the basis of the particle swarm algorithm. During the particle update process, two groups of SKU arrangements are randomly selected regularly for crossover operation. By exchanging the positions of some SKUs, new storage location combinations are generated to broaden the search space; at the same time, a random mutation mechanism is introduced to randomly exchange the arrangements of SKUs on the storage locations. The genetic algorithm formula is:

[0056]

[0057] Among them, S1 and S2 are the SKU arrangements of two particles, S is the current SKU arrangement, and p c is the crossover rate, controlling the frequency of the crossover operation, taking 0.7 - 0.9, and p m is the mutation rate, controlling the frequency of the mutation operation, usually taking 0.1 - 0.2, and t is the interval generation number of the genetic operation, taking 5 - 20.

[0058] Furthermore, in step 6, the dynamic adjustment of the temperature parameter is specifically as follows:

[0059] The temperature T(t) of the t-th generation is as follows:

[0060] T(t) = T0·α t ,

[0061] where T0 is the initial temperature, α is the temperature reduction coefficient, taking 0.9 ≤ α ≤ 0.99, and t is the current iteration number; when the initial temperature is higher, the simulated annealing algorithm is more likely to accept a worse solution, increasing the probability of jumping out of the local optimal solution. As the iteration number increases, the temperature gradually decreases, the acceptance probability of the worse solution decreases, and it converges to the global optimal solution.

[0062] Furthermore, the mechanism for accepting a worse solution is specifically as follows:

[0063] The acceptance probability of a worse solution is based on the Metropolis criterion:

[0064]

[0065] where P is the probability of accepting a worse solution, ΔE = E new -E current is the fitness difference between the new solution and the current solution. When ΔE < 0, it means the new solution is better than the current solution, and the new solution must be accepted; when ΔE > 0, it means the new solution is worse than the current solution, and the worse solution is accepted with probability P; when the temperature is high, P is large, and it is easier to accept a worse solution and jump out of the local optimum; when the temperature is low, P is small, gradually converging and stabilizing near the global optimum.

[0066] The beneficial effects of the present invention include: By combining the particle swarm algorithm with the genetic algorithm and the simulated annealing algorithm, the storage location distribution is optimized, significantly improving the storage efficiency and the material flow speed.

[0067] First, this method designs a warehouse space model based on the Manhattan distance. By performing a grid modeling on the three-dimensional coordinates of the storage space, the Manhattan distances from the outbound and inbound positions to each storage location are calculated, thereby quantifying the material flow cost in the warehouse.

[0068] Then, this method introduces an inbound and outbound frequency prediction model based on the LSTM long short-term memory network. By performing a time series modeling on the historical inbound and outbound frequency data, the inbound and outbound frequency distribution of SKUs in the future period is predicted, and random perturbations and stretching coefficients are added to the prediction results to expand the frequency distribution range and prevent the repetition of the predicted frequencies, thereby enhancing the generalization ability of the model.

[0069] Finally, the present method uses an improved particle swarm optimization algorithm for warehouse location planning. The particle swarm optimization algorithm achieves a dynamic balance between search and convergence capabilities by adaptively adjusting the inertia weight; the genetic algorithm introduces crossover and mutation operations during the particle update process to enhance population diversity and prevent local optima; the simulated annealing algorithm jumps out of the local optimal solution and improves the global search ability by introducing a temperature parameter and an inferior solution acceptance mechanism. The overall fitness function comprehensively considers factors such as inbound and outbound frequencies, Manhattan distance of storage locations, cargo weight, and volume during the optimization process. At the same time, a frequency dispersion penalty term is introduced to prevent high-frequency SKUs from being concentrated in stacks, optimize the overall frequency distribution, and further improve the warehousing efficiency.

[0070] This method can visually display the optimized warehouse layout, with a more reasonable distribution of storage locations. High-frequency SKUs are placed close to the entrances and exits, effectively reducing the length of the material flow path and time consumption, and significantly improving the operation efficiency and management ability of the intelligent warehousing system. Description of the Drawings

[0071] Figure 1 is a flowchart of the intelligent warehouse location planning and optimization method according to an embodiment of the present invention;

[0072] Figure 2 is a contour Boolean matrix diagram of the storage location and cargo characteristics according to an embodiment of the present invention;

[0073] Figure 3 is a schematic diagram of the LSTM inbound and outbound frequency prediction model according to an embodiment of the present invention;

[0074] Figure 4 is a comparison diagram of the data frequency before and after expansion according to an embodiment of the present invention;

[0075] Figure 5 is a diagram of the optimized result of the intelligent warehouse location planning according to an embodiment of the present invention. Detailed Embodiments

[0076] In order to make the objectives, technical solutions, and technical effects of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings of the specification and embodiments.

[0077] As Figure 1 shown, an embodiment of the present invention discloses an intelligent warehouse location planning and optimization method based on a hybrid particle swarm optimization algorithm, including:

[0078] Step 1: Establish a warehouse space model based on the Manhattan distance and design an overall fitness function.

[0079] Specifically, a warehouse space model based on Manhattan distance is designed. First, a three-dimensional coordinate grid model of the storage space is established. The storage area is divided into regular three-dimensional grid cells composed of length, width, and height. The Manhattan distances from the outbound and inbound positions to each storage location are calculated, and a height penalty term is introduced to quantify the transfer cost of materials in the warehouse.

[0080] The formula for calculating the Manhattan distance is as follows:

[0081] D = |x - x0| + |y - y0| + |h - h0|,

[0082] where (x0, y0, h0) is the position of the outbound or inbound port, (x, y, h) is the three-dimensional coordinates of the target storage location, and D is the Manhattan distance between the entrance / exit and the target storage location. This distance metric can accurately reflect the actual number of grid cells that materials need to cross during the storage and retrieval process in a regular storage grid, thus effectively quantifying the transfer path and handling cost of materials in the warehouse.

[0083] Since in the actual storage environment, the difficulty and time consumption of accessing materials on high-level shelves are relatively large, when the model calculates the Manhattan distance, it considers the multi-level shelf structure of the warehouse, incorporates the height dimension into the distance calculation, and introduces a height penalty term:

[0084] D′ = D + β·h,

[0085] where β is the height penalty coefficient and h is the height of the storage location.

[0086] Meanwhile, to prevent the accumulation of high-frequency SKUs in local areas, the embodiment of the present invention introduces a frequency dispersion penalty term, and the calculation formula is:

[0087]

[0088] where λ is the penalty coefficient used to control the influence weight of the penalty term in the fitness function, N is the number of SKUs, α is the sensitivity coefficient of frequency difference to control the convergence speed of the penalty term, |f i - f j | is the difference in outbound frequencies between SKUi and SKUj, and D(i, j) is the Manhattan distance between SKUi and SKUj. When the inbound and outbound frequencies of SKU i and SKU j are close, that is, |f i - f j | is small, the value of the penalty term is large, prompting the algorithm to store them dispersedly. When the difference in inbound and outbound frequencies between SKUi and SKUj is large, that is, |f i - f j|Larger, the penalty term tends to 0, and SKUs with large frequency differences can be relatively concentrated. When the bin distance D(i,j) increases, the penalty term decreases, encouraging SKUs with similar frequencies to be distributed farther apart in space. The penalty term increases rapidly as the frequency difference decreases and the distance decreases, thus effectively preventing SKUs with similar frequencies from accumulating locally and promoting a more balanced distribution of SKUs in the bin space.

[0089] Design the overall fitness function in warehouse optimization, and the overall fitness function is:

[0090]

[0091] Among them, f out,i and f in,i are the inbound and outbound frequencies of SKUi, D out,i and D in,i are the Manhattan distances of the SKU to the entrance and exit, w i and v i are the weight and volume of the SKU, h is the floor height of the SKU in the warehouse space, Penalty is the frequency dispersion penalty term, which prevents the accumulation of high-frequency SKUs, optimizes the overall frequency distribution, and further improves the warehouse efficiency. This function is used as an evaluation index to measure the quality of the SKU bin distribution. As Figure 2 shown is the contour Boolean matrix diagram of the bin location and cargo characteristics.

[0092] Step 2: As Figure 3 shown, establish an inbound and outbound frequency prediction model based on the LSTM long short-term memory network.

[0093] By performing time series modeling on the historical inbound and outbound frequency data, predict the change trend of the SKU inbound and outbound frequencies in the future period.

[0094] In the data preprocessing stage, normalize the multi-dimensional historical frequency data to eliminate the numerical differences between the frequencies of different SKUs and prevent the problem of gradient explosion or disappearance during model training. The min-max normalization formula is:

[0095]

[0096] Among them, f' i,j and g' i,j are the normalized inbound and outbound frequencies respectively, f i,j and g i,j are the original frequencies of the i-th time step and the j-th SKU respectively, min(f j ) and max(f jare the minimum and maximum outbound frequencies of the j-th SKU at all time steps. After normalization, all frequency data is compressed into the range [0, 1] to prevent gradient explosion or vanishing problems caused by large numerical differences during model training.

[0097] Next, the historical frequency data is divided into time windows to construct a sliding time series dataset:

[0098] X (k) = [f' k , f' k+1 , …, f' k+τ-1 ,

[0099] y (k) = f' k+τ ,

[0100] where X (k) is the input sequence of the k-th time window, with dimension τ×N, and y (k) is the predicted frequency at the subsequent moment of the corresponding time window.

[0101] The time dependence and long-term trend of historical frequency changes are learned through an LSTM network. The model adopts a multi-layer LSTM structure, where the hidden layer is used to extract the time features of frequency changes, and the output layer uses a linear activation layer to output the predicted values of the inbound and outbound frequencies of the SKU at future moments. The predicted future frequency distribution is:

[0102]

[0103] where W out is the weight of the last layer, b out is the bias, and h t is the hidden state at the current time step.

[0104] The model parameters are optimized through adaptive learning rate adjustment and the mean squared error (MSE) loss function. The formula is:

[0105]

[0106] where is the predicted frequency, and f i is the true frequency. The optimization uses the Adam adaptive learning rate algorithm for parameter adjustment to prevent getting stuck in local optimal solutions and ensure that the model can effectively capture the frequency fluctuation characteristics during the fitting process and improve the prediction accuracy.

[0107] Step 3: Perform random perturbation and stretching on the model prediction results to expand the frequency distribution range.

[0108] Specifically, after predicting the inbound and outbound frequencies, to prevent the future frequency distribution generated by the model from being too smooth or having too high a frequency repetition rate, the prediction results are further processed by random perturbation and stretching to increase the difference and diversity of the frequency distribution.

[0109] First, random perturbation is introduced by superimposing Gaussian noise on the prediction results to simulate the possible volatility and uncertainty of future frequencies, thus avoiding too regular a frequency distribution. Subsequently, a stretching coefficient is introduced to expand the frequency distribution range, and some frequency values are moderately amplified or reduced to increase the difference and diversity of the frequency distribution. After combining the random perturbation and stretching processing, the future frequency distribution is:

[0110]

[0111] where 1+∈ is the random fluctuation coefficient and S is the stretching coefficient.

[0112] As Figure 4 shown, the comparison of data frequency before and after expansion.

[0113] Step 4: Based on the adaptive particle swarm optimization algorithm, by adaptively adjusting the inertia weight, the dynamic search and convergence balance of the storage location optimization are achieved.

[0114] Specifically, the particle swarm optimization algorithm simulates the search behavior of particles in the solution space, continuously adjusts their positions and velocities to seek the optimal storage location distribution. The velocity update formula and position update formula of the particles are respectively:

[0115] v i (t + 1) = w(t)·v i (t) + c1·r1·(p best,i -x i (t)) + c2·r2·(g best -x i (t)),

[0116] x i (t + 1) = x i (t) + v i (t + 1),

[0117] where x i (t) is the storage location distribution of the i-th particle at the t-th iteration, v i (t) is the velocity of the i-th particle, p best,i is the historical optimal solution of the i-th particle, g best is the global optimal solution in the population history, w(t) is the inertia weight, c1 and c2 are learning factors representing acceleration coefficients, and r1 and r2 are random numbers in the range [0,1][0,1][0,1].

[0118] An adaptive inertia weight mechanism is designed to enable the particles to dynamically adjust their exploration and convergence capabilities during the search process. The inertia weight formula is as follows:

[0119]

[0120] where w max is the initial inertia weight, usually set to 0.90 to enhance the exploration ability. w min is the final inertia weight, usually set to 0.40 to enhance the convergence ability. t is the current iteration number, and T is the maximum iteration number. At the beginning of the iteration, a larger inertia weight is adopted to enhance the global exploration, so as to conduct a more extensive search in the storage location space and prevent falling into the local optimum. At the later stage of the iteration, the inertia weight is gradually reduced to accelerate the convergence speed of the algorithm, so as to finely adjust the storage location distribution and ensure the optimized access efficiency of SKUs.

[0121] In addition, the velocity update of the particles adopts a probability mapping mechanism, which maps the velocity of the particles into a probability through the Sigmoid function:

[0122]

[0123] where P i,j is the jump probability of particle i in dimension j, and v i,j (t) is the velocity of the i-th particle in the j-th dimension, to introduce an appropriate amount of randomness, prevent the search process from being too rigid, and help jump out of the local optimal solution.

[0124] Step 5: Based on the genetic algorithm, crossover and mutation operations are introduced during the particle update process to enhance the diversity of the storage location distribution and prevent local optimality.

[0125] Specifically, to further improve the diversity and flexibility of the storage location optimization, the present invention integrates the genetic algorithm GA on the basis of the particle swarm algorithm. The genetic algorithm introduces crossover and mutation operations during the storage location optimization process. Specifically, during the particle update process, two groups of SKU arrangements are randomly selected at regular intervals for crossover operations. By exchanging the positions of some SKUs, new storage location combinations are generated to broaden the search space. At the same time, the algorithm introduces a random mutation mechanism to randomly exchange the arrangements of SKUs on the storage locations with a certain probability. This mutation operation breaks the limitation of the fixed solution and introduces new possibilities for the storage location layout, thereby preventing the population from falling into the local optimal solution during the search process. The genetic algorithm formula is as follows:

[0126]

[0127] where S1 and S2 are the SKU arrangements of two particles, S is the current SKU arrangement, and p cis the crossover rate, which controls the frequency of the crossover operation. It is usually taken as 0.7 - 0.9, p m is the mutation rate, which controls the frequency of the mutation operation. It is usually taken as 0.1 - 0.2, and t is the interval generation of the genetic operation, which is usually taken as 5 - 20. The crossover and mutation operations effectively break the local optimal limit, enhance the diversity of the solution space, maintain a high solution space exploration ability during the storage location optimization process, and further improve the access efficiency of SKUs and the balance of the overall distribution.

[0128] Step 6: Based on the simulated annealing algorithm, by introducing the temperature parameter, jump out of the local optimal solution, enhance the global storage location optimization ability, and break the frequency accumulation effect.

[0129] Specifically, introduce the simulated annealing algorithm into the hybrid optimization framework of the particle swarm and genetic algorithms. By introducing the temperature parameter and the inferior solution acceptance mechanism, improve the ability of the algorithm to jump out of the local optimal solution, prevent high-frequency SKUs from accumulating near the entrance and exit, and promote the more balanced distribution of SKUs in space. The core formula of the simulated annealing algorithm is the dynamic adjustment of the temperature parameter. The temperature T(t) in the t-th generation is:

[0130] T(t) = T0·α t ,

[0131] where, T0 is the initial temperature, α is the temperature reduction coefficient, usually taken as 0.9 ≤ α ≤ 0.99, and t is the current iteration number. When the initial temperature is higher, the algorithm is more likely to accept inferior solutions, increasing the probability of jumping out of the local optimal solution. As the iteration number increases, the temperature gradually decreases, and the inferior solution acceptance probability decreases, converging to the global optimal solution. For the selection of inferior solutions, the acceptance probability of inferior solutions is based on the Metropolis criterion:

[0132]

[0133] where, P is the probability of accepting the inferior solution, ΔE = E new -E current is the fitness difference between the new solution and the current solution, and T(t) is the current temperature. When the new solution is better than the current solution (ΔE < 0), the new solution must be accepted; when the new solution is inferior to the current solution (ΔE > 0), accept the inferior solution with probability P. When the temperature is high, P is large, and it is easier to accept inferior solutions and jump out of the local optimum. When the temperature is low, P is small, gradually converging and stabilizing near the global optimum.

[0134] Finally, visually display the optimized warehouse layout. As Figure 5 shown, the results of the storage location planning optimization of the present invention show that the method of the present invention has obtained good results in establishing a warehouse model, predicting future inventory data, and intelligent storage location allocation in intelligent storage location optimization.

[0135] The above are only the preferred embodiments of the present invention and do not impose any form of limitation on the present invention. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligent warehouse location planning optimization method based on a hybrid particle swarm algorithm, characterized in that, Including: Step 1: Establish a warehouse space model based on Manhattan distance and design an overall fitness function; Step 2: Establish an inbound and outbound frequency prediction model based on the LSTM long short-term memory network; Step 3: Perform random perturbation and stretching processing on the prediction results of the inbound and outbound frequency prediction model to expand the frequency distribution range; Step 4: Based on the particle swarm algorithm, achieve dynamic search and convergence balance of storage location optimization by adaptively adjusting the inertia weight; Step 5: Based on the genetic algorithm, introduce crossover and mutation operations during the particle update process to enhance the diversity of storage location distribution; Step 6: Based on the simulated annealing algorithm, introduce a temperature parameter and a bad solution acceptance mechanism to jump out of the local optimal solution and improve the global storage location optimization ability.

2. The intelligent warehouse location planning and optimization method according to claim 1, wherein In Step 1, a grid-based modeling of the three-dimensional coordinates of the storage space is carried out, the Manhattan distances from the outbound and inbound positions to each storage location are calculated, and a height penalty term is introduced to quantify the transfer cost of materials in the warehouse; The formula for calculating the Manhattan distance is: D = |x - x0| + |y - y0| + |h - h0|, where (x0, y0, h0) is the position of the outbound or inbound port, (x, y, h) is the three-dimensional coordinates of the target storage location, and D is the Manhattan distance between the entrance / exit and the target storage location; When calculating the Manhattan distance in the warehouse space model, a height penalty term is introduced: D′ = D + β·h, where β is the height penalty coefficient and h is the height where the storage location is located; A frequency dispersion penalty term is also introduced, and the calculation formula is: Among them, λ is the penalty coefficient, which is used to control the influence weight of the penalty term in the overall fitness function. N is the number of stock keeping units (SKUs). α is the sensitivity coefficient of frequency difference, which controls the convergence speed of the penalty term. |f i -f j | is the difference in outbound frequencies between SKUi and SKUj, and D(i, j) is the Manhattan distance between SKUi and SKUj. When the inbound and outbound frequencies of SKUi and SKUj are close, the value of the penalty term is large, prompting the algorithm to store them separately. When the difference in inbound and outbound frequencies between SKUi and SKUj is large, the penalty term tends to 0, and the SKUs with large frequency differences are relatively concentrated. When the bin distance D(i, j) increases, the penalty term decreases, encouraging SKUs with similar frequencies to be distributed farther apart in space. The penalty term increases rapidly as the frequency difference and distance decrease.

3. The intelligent warehouse location planning and optimization method according to claim 2, characterized in that In Step 1, the overall fitness function is: Among them, f out,i and f in,i are the inbound and outbound frequencies of SKU i, D out,i and D in,i are the Manhattan distances from the SKU to the entrance and exit, w i and v i are the weight and volume of the SKU, and h is the floor height of the SKU in the storage space.

4. The intelligent warehouse location planning and optimization method according to claim 1, wherein The specific content of Step 2 includes: First, in the data preprocessing stage, using multi-dimensional historical frequency data, the historical frequency data is normalized, and the normalization formula is: where f′ i,j and g′ i,j are the normalized inbound and outbound frequencies respectively, and f i,j and g i,j are the original frequencies of the i-th time step and the j-th SKU respectively. min(f j ) and max(f j ) are the minimum and maximum outbound frequencies of the j-th stock-keeping unit (SKU) over all time steps. After normalization, all frequency data is compressed into the interval [0, 1]. Next, the historical frequency data is divided into time windows to construct a sliding time series data set: Among them, X (k) is the input sequence of the k-th time window, with a dimension of τ×N, and y (k) is the predicted frequency at the subsequent moment corresponding to the time window; An inbound and outbound prediction model is established using a multi-layer LSTM structure, where the hidden layer is used to extract the time features of frequency changes, and the output layer uses a linear activation layer to output the predicted values of the inbound and outbound frequencies of the SKU at future times, and the predicted future frequency distribution is as follows: Among them, W out is the weight of the last layer, b out is the bias, and h t is the hidden state at the current time step; The model parameters are optimized through adaptive learning rate adjustment and mean square error loss function, and the formula is: Among them, is the predicted frequency, f i is the true frequency; The optimization of the inbound and outbound prediction model uses the Adam adaptive learning rate algorithm for parameter adjustment.

5. The intelligent warehouse location planning and optimization method according to claim 1, characterized in that The specific content of Step 3 includes: First, random perturbation is introduced by superimposing Gaussian noise on the prediction results to simulate the possible volatility and uncertainty of future frequencies, thereby avoiding overly regular frequency distribution; Subsequently, a stretching coefficient is introduced to expand the frequency distribution range, and some frequency values are moderately amplified or reduced to increase the difference and diversity of the frequency distribution; After combining the random perturbation and stretching processing, the future frequency distribution is: where 1 + ∈ is the random fluctuation coefficient and S is the stretching coefficient.

6. The intelligent warehouse location planning and optimization method according to claim 1, wherein, In step 4, the particle swarm algorithm simulates the search behavior of particles in the solution space, continuously updates and adjusts their positions and velocities to seek the optimal storage location distribution. The velocity update formula and position update formula of the particles are respectively: v i (t + 1) = w(t)·v i (t) + c1·r1·(p best,i - x i (t)) + c2·r2·(g best - x i (t)), x i (t + 1) = x i (t) + v i (t + 1), where x i (t) is the library location distribution of the i-th particle at the t-th iteration, v i (t) is the velocity of the i-th particle, p best,i is the optimal solution in the history of the i-th particle, g best is the global optimal solution in the history of the population, w(t) is the inertia weight, c1 and c2 are learning factors representing the acceleration coefficients, and r1 and r2 are random numbers in the range [0,1][0,1][0,1].

7. The intelligent warehouse location planning and optimization method according to claim 6, characterized in that, The velocity update of the particle adopts a probability mapping mechanism, and the velocity of the particle is mapped to a probability through the Sigmoid function: Among them, P i,j is the jump probability of particle i in dimension j, and v i,j (t) is the velocity of the i-th particle in the j-th dimension.

8. The intelligent warehouse location planning and optimization method according to claim 6, wherein In Step 4, an adaptive inertia weight mechanism is introduced, and the inertia weight formula is: Among them, w max is the initial inertia weight, set to 0.90 to enhance the exploration ability, and w min is the final inertia weight, set to 0.40 to enhance the convergence ability; t is the current iteration number, and T is the maximum iteration number; in the initial stage of iteration, a larger inertia weight is adopted to enhance the global exploration, so as to conduct a wider search in the storage location space and prevent falling into the local optimum; in the later stage of iteration, the inertia weight is gradually reduced to accelerate the convergence speed of the algorithm, so as to finely adjust the storage location distribution and ensure the optimization of the access efficiency of the stock keeping unit (SKU).

9. The intelligent warehouse location planning and optimization method according to claim 6, characterized in that In step 5, the genetic algorithm GA is fused on the basis of the particle swarm algorithm. During the particle update process, two sets of stock - keeping unit (SKU) arrangements are randomly selected regularly for crossover operations. By exchanging the positions of some SKUs, new storage location combinations are generated to broaden the search space. At the same time, a random mutation mechanism is introduced to randomly exchange the arrangements of SKUs on the storage locations. The genetic algorithm formula is as follows: Among them, S1 and S2 are the SKU arrangements of two particles, S is the current SKU arrangement, and p c is the crossover rate, which controls the frequency of the crossover operation, taking values from 0.7 to 0.9, and p m is the mutation rate, which controls the frequency of the mutation operation, usually taking values from 0.1 to 0.2, and t is the interval generation number of the genetic operation, taking values from 5 to 20.

10. The intelligent warehouse location planning and optimization method according to claim 1, characterized in that, In step 6, the dynamic adjustment of the temperature parameter is specifically as follows: The temperature T(t) in the t - th generation is: T(t) = T0·α t , where T0 is the initial temperature, α is the temperature - decreasing coefficient, with 0.9 ≤ α ≤ 0.99, and t is the current iteration number. When the initial temperature is higher, the simulated annealing algorithm is more likely to accept inferior solutions, increasing the probability of jumping out of the local optimal solution. As the number of iterations increases, the temperature gradually decreases, and the probability of accepting inferior solutions decreases, converging to the global optimal solution.

11. The intelligent warehouse location planning and optimization method according to claim 10, wherein The inferior - solution acceptance mechanism is specifically as follows: The acceptance probability of the inferior solution is based on the Metropolis criterion: where P is the probability of accepting a worse solution, and ΔE = E new - E current is the fitness difference between the new solution and the current solution. When ΔE < 0, it means the new solution is better than the current solution, and thus the new solution must be accepted. When ΔE > 0, it means the new solution is worse than the current solution, and then the worse solution is accepted with probability P. When the temperature is high, P is large, making it easier to accept a worse solution and jump out of the local optimum. When the temperature is low, P is small, gradually converging and stabilizing near the global optimum.

Images