Big data driven logistics demand prediction platform

Abstract

The invention discloses a logistics demand prediction platform driven by big data, and relates to the technical field of logistics intelligent scheduling. The platform comprises a dynamic feature perception module, a space-time prediction engine, a simulation deduction and optimization module and an online learning and feedback module. The dynamic feature sensing module fuses multi-source heterogeneous logistics time sequence data, constructs and updates a dynamic logistics knowledge graph, and extracts space-time dynamic features; the space-time prediction engine captures the space-time dependence of the features through the hierarchical space-time attention network, and outputs a multi-period demand prediction sequence; the simulation deduction and optimization module simulates various external disturbance scenes based on a discrete event simulation model, and generates a priority-sorted elastic scheduling plan through Monte Carlo simulation and multi-objective optimization; and the online learning and feedback module adaptively optimizes model parameters and feature extraction rules through closed-loop iteration. According to the method, the problems of insufficient multi-source data fusion, inaccurate space-time dependence capture and weak anti-disturbance capability in traditional logistics prediction are solved.

Description

Technical Field

[0001] This invention relates to the field of logistics big data processing and intelligent scheduling technology, specifically to a big data-driven logistics demand forecasting platform that integrates spatiotemporal prediction and simulation optimization. Background Technology

[0002] With the rapid development of e-commerce and instant logistics, the scale of logistics networks is constantly expanding, and the connections between nodes are becoming increasingly complex. Logistics demand exhibits characteristics such as strong spatiotemporal dependence, suddenness, and uncertainty. Existing logistics demand forecasting technologies have the following shortcomings:

[0003] The integration of multi-source heterogeneous data (order data, GPS trajectory, weather data, road network status, etc.) is insufficient, making it difficult to comprehensively represent the dynamics of the logistics network;

[0004] Predictive models have limited ability to capture spatiotemporal dependencies, and traditional time series models (ARIMA, LSTM) cannot effectively characterize the dynamic relationships between nodes and the long-term and short-term time series patterns.

[0005] Lacking robustness to external disturbances such as extreme weather and node failures, the dispatch plans are mostly static and difficult to adapt to uncertain scenarios;

[0006] The lack of a closed-loop optimization mechanism in the prediction model and scheduling strategy makes it impossible to dynamically adjust based on the actual execution results, leading to a decline in long-term prediction accuracy and scheduling efficiency. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a big data-driven logistics demand forecasting platform to solve the aforementioned problems.

[0009] (II) Technical Solution

[0010] To achieve the above objectives, this invention provides a big data-driven logistics demand forecasting platform, the specific technical solution of which is as follows:

[0011] Platform Overall Architecture

[0012] The platform includes a dynamic feature perception module, a spatiotemporal prediction engine, a simulation and optimization module, and an online learning and feedback module. These modules are connected sequentially to form a data flow and optimization closed loop, as detailed below:

[0013] The dynamic feature perception module accesses and integrates multi-source heterogeneous logistics time-series data, builds and updates a dynamic logistics knowledge graph (dynamic network model) in real time, and extracts spatiotemporal dynamic features.

[0014] The dynamic feature perception module supports access to multi-source heterogeneous data such as order data, GPS trajectory data, warehouse inventory data, road network status data, weather data, and promotional activity data. After data cleaning (removing missing values ​​and outliers), standardization (Min-Max mapping to the [0,1] interval), and spatiotemporal alignment (based on timestamps accurate to the second and geocoding latitude and longitude matching), a time series dataset in a unified format is formed.

[0015] Construction and updating of dynamic logistics knowledge graph:

[0016] Entity definition: The core entities are logistics nodes (warehouses, distribution centers, stations), transportation routes, and orders. Entity attributes include node inventory, processing capacity, route distance, and order timeliness requirements.

[0017] Edge definition: The core edges are "node-route" association (transportation relationship) and "order-node" association (send and receive relationship), and the edge attribute is the association strength;

[0018] Update mechanism: A hybrid mechanism of time window sliding (configurable from 15 minutes to 1 hour, default 30 minutes) and key event triggering (order volume exceeding twice the historical average, extreme weather warning) is adopted to update the entity attributes and edge association strength of the knowledge graph in real time;

[0019] Spatiotemporal dynamic feature extraction: Two types of core features are extracted from the dynamic logistics knowledge graph, which completely correspond to claim 2:

[0020] Node state vector: Represents the state of each logistics node at time t, with a dimension of N×D (N is the number of nodes, D is the feature dimension, and the default D=12). Features include node inventory level, processing capacity utilization, number of orders in transit, historical demand average, weather impact coefficient, road network congestion index, etc.

[0021] The association matrix represents the association strength between nodes at time t. It has N×N dimensions and is calculated by weighting the following factors: transportation distance (0.3), transportation timeliness (0.25), capacity saturation (0.2), historical collaboration frequency (0.15), and current order flow direction (0.1). The formula is as follows: (d ij Let T be the distance between nodes ij. ij For timely delivery, S ij For capacity saturation, F ij O is the normalized value of historical collaboration frequency. ij (This represents the current percentage of order flow).

[0022] Furthermore, the spatiotemporal prediction engine is unidirectionally connected to the dynamic feature perception module (receiving spatiotemporal dynamic features) and outputs a demand prediction sequence for multiple future time periods.

[0023] The spatiotemporal prediction engine incorporates a hierarchical spatiotemporal attention network (a prediction model based on an attention mechanism), which fully corresponds to claims 3 and 4, and includes:

[0024] Spatial Attention Unit (Graph Attention Subnetwork): Taking the node state vector and incidence matrix as input, it calculates the dynamic spatial dependencies between nodes. The specific process is as follows:

[0025] S1, regarding the node state vector h i Perform a linear transformation: h i ′=W h ·h i (W h The weight matrix has dimensions D×H, where H is the dimension of the hidden layer (default H=64).

[0026] S2. Calculate the attention scores for nodes i and j: e ij =LeakyReLU(W a [h i ′||h j ′]·a)(W a (where is the weight matrix, dimension 2H×1, and a is the attention coefficient vector, dimension 1×1);

[0027] S3, Based on the correlation matrix M ij Filtering invalid associations (M) ij time e ij =0), and the attention weights are obtained by softmax normalization: α ij =softmax j (e ij );

[0028] S4, Weighted Aggregation of Neighbor Node Features: (σ is the ReLU activation function), generating the hidden state of the node containing spatial context information;

[0029] Temporal attention unit (temporal attention subnetwork): based on the sequence of historical hidden states of each node. (T is the length of the historical time series, default T=24) is the input, adaptively capturing time series patterns:

[0030] S1. Calculate the attention weight of time t' relative to the current time t: (MLP is a 2-layer perceptron with an output dimension of 1);

[0031] S2, Temporal Feature Aggregation:

[0032] Predicted output unit: fusing spatial and temporal features The input consists of a fully connected layer (2H×K dimensions, where K is the number of prediction time periods; the default K=12, corresponding to the next 12 hours, with one prediction value per hour) and a temporal decoding layer (based on GRU), which decodes and outputs the demand prediction sequence. ( (The demand forecast for node i).

[0033] Furthermore, the simulation and optimization module is connected to the spatiotemporal prediction engine and the dynamic feature perception module respectively, and receives the demand prediction sequence and the current network state.

[0034] Specific implementation steps:

[0035] Step 1: Construct a discrete event simulation model for the logistics network: Based on a dynamic logistics knowledge graph, integrate node processing flow (inbound → sorting → outbound, with processing time for each stage following a log-normal distribution), route transportation flow (loading → en route → unloading, with en route time dynamically adjusted based on road network status), and resource constraints (number of vehicles, storage capacity, and personnel configuration as hard constraints), using the Python SimPy framework.

[0036] Step 2: Generate basic scheduling plans: Using the demand forecast sequence as the basic input, combine it with 5 predefined scheduling rules (corresponding to M = 5 basic plans):

[0037] Rule 1 (Cost Priority): Select the resource allocation scheme with the lowest combined transportation and warehousing costs;

[0038] Rule 2 (Time-sensitive priority): Select the route planning scheme with the shortest delivery time;

[0039] Rule 3 (Balanced): Balancing capacity utilization with timeliness deviation;

[0040] Rule 4 (Emergency Priority): Reserve 10% of capacity to cope with sudden demand;

[0041] Rule 5 (Regional Collaboration): Adjacent nodes share transportation resources;

[0042] Step 3: External disturbance scene injection and Monte Carlo simulation:

[0043] Three predefined disturbance scenarios (user extension supported):

[0044] Extreme weather: Traffic efficiency on core routes will decrease by 30%-50% (random value), lasting for 2-4 hours;

[0045] Node failure: The processing capacity of the regional distribution center decreases by 50%-100% (random value), lasting for 1-3 hours;

[0046] Demand surge: Localized demand increases by 100%-200% (random value), lasting for 1-2 hours;

[0047] Perturbation Combinations: Each basic pre-set is injected with random combinations of 3 types of scenarios (27 combinations in total), and 1000-5000 Monte Carlo simulations are performed (2000 by default), recording the performance data of each simulation;

[0048] Step 4: Multi-objective optimization and contingency plan ranking:

[0049] Performance metrics: Define 3 core metrics (consistent with the technical solution):

[0050] Delivery time deviation:

[0051] Dispatch cost: C = Transportation cost + Warehousing cost + Emergency cost (unit: yuan);

[0052] Order fulfillment rate:

[0053] Multi-objective optimization: The NSGA-III algorithm is adopted, with the objective functions being min(E,C) and max(R), and the constraint condition being R≥90%.

[0054] Plan ranking: After generating the Pareto optimal solution set, calculate the comprehensive score according to "timeliness deviation weight 0.4 + cost weight 0.3 + satisfaction rate weight 0.3", and output the Top 3 flexible scheduling plan set (sorted by priority).

[0055] Furthermore, the online learning and feedback module forms a closed loop with other modules, enabling the platform to adapt and iterate.

[0056] The online learning and feedback module collects three types of data in real time:

[0057] i. Prediction error:

[0058] ii. Deviation in execution effect: ΔE = |E 实际 -E 仿真 |、ΔC=|C 实际 -C 仿真 |、ΔR=|R 实际 -R 仿真 |;

[0059] iii. Latest logistics data: new orders, real-time road network status, changes in node inventory, etc.;

[0060] Adaptive optimization of the online learning and feedback module:

[0061] Model parameter optimization: Based on prediction error and new data, mini-batch gradient descent (batch size 32, learning rate 0.001) is used to update the weight parameters (W) of the hierarchical spatiotemporal attention network. h W a a) Dynamically adjust the historical time series length T (range 12-48) and the prediction step size K (range 6-24);

[0062] Feature extraction rule optimization: Based on the execution effect deviations ΔE and ΔC, adjust the feature weights of the node state vector (e.g., increase the weight of the weather impact coefficient when the timeliness deviation is too large) and the calculation coefficients of the correlation matrix (e.g., increase the weight of the transportation distance when the cost deviation is too large).

[0063] Knowledge graph update mechanism optimization: dynamically adjust the time window size (15 minutes - 1 hour) or the threshold for triggering key events (such as adjusting the threshold for a sudden increase in order volume from 2 times the historical average to 1.8 times);

[0064] Iteration mechanism: The default optimization cycle is 24 hours. When the prediction error... i When the mean of (t) exceeds 15%, immediate optimization is triggered. The optimization process uses incremental learning and does not interrupt the platform operation.

[0065] (II) Supplementary Technical Features

[0066] The dynamic logistics knowledge graph supports the storage and real-time updates of tens of millions of nodes and edges, and uses HBase distributed storage.

[0067] The training process of the hierarchical spatiotemporal attention network adopts an early stopping mechanism (the training stops if the validation set error does not decrease for 5 consecutive rounds) to avoid overfitting;

[0068] The time step of the discrete event simulation model is 1 minute to ensure simulation accuracy.

[0069] The optimization process of the online learning and feedback module records historical parameters and supports a rollback mechanism.

[0070] Compared with existing technologies, this invention provides a big data-driven logistics demand forecasting platform, which has the following beneficial effects:

[0071] This invention achieves comprehensive dynamic characterization of logistics networks through deep fusion of multi-source heterogeneous data. It utilizes a dynamic feature perception module to access various heterogeneous data sources, including orders, GPS trajectories, warehouse inventory, road network status, weather, and promotional activities. After meticulous processing such as cleaning, standardization, and spatiotemporal alignment, a unified time-series dataset is formed. Furthermore, a dynamic logistics knowledge graph is constructed based on core entities such as logistics nodes, transportation routes, and orders. A hybrid mechanism of sliding time windows and key event triggers updates entity attributes and association strength in real time, solving the problem of insufficient data fusion in traditional technologies. The extracted node state vectors (12-dimensional core features) and association matrices (multi-dimensional weighted calculations) can comprehensively and dynamically depict the node states and inter-node relationships of the logistics network, providing accurate data support for subsequent prediction and scheduling.

[0072] This invention innovatively employs a hierarchical spatiotemporal attention network as the core of prediction, accurately capturing spatiotemporal dependencies and significantly improving demand forecasting accuracy. It dynamically calculates spatial dependency weights between nodes through a graph attention sub-network (spatial attention unit), filters invalid associations by fusing an association matrix, and generates features containing spatial context. Then, a temporal attention sub-network (temporal attention unit) adaptively captures the temporal evolution patterns of historical states, achieving deep integration of spatial and temporal features. Compared to the shortcomings of traditional ARIMA and LSTM models in effectively characterizing dynamic spatial relationships and long- and short-term temporal patterns, the demand forecast sequences for the next 12 hours (configurable from 6 to 24 hours) output by this platform have a mean absolute percentage error (MAPE) that is more than 40% lower than traditional models, significantly improving prediction accuracy and timeliness.

[0073] To enhance resilience against disturbances and generate adaptive flexible scheduling plans, this invention addresses uncertainties such as extreme weather, node failures, and sudden demand changes. It constructs a discrete event simulation model of the logistics network through simulation and optimization modules, injecting 27 combinations of disturbance scenarios into 5 basic scheduling plans and executing 2000 (configurable 1000-5000) Monte Carlo simulations to comprehensively evaluate the plan's performance under complex disturbances. Combining the NSGA-III multi-objective optimization algorithm, with delivery timeliness deviation, scheduling cost, and order fulfillment rate as core indicators (constraining order fulfillment rate ≥90%), it generates a top 3 set of flexible scheduling plans ranked by priority. This solution overcomes the limitations of traditional static scheduling plans, adaptively adjusting resource allocation and path planning strategies according to different disturbance scenarios. Under extreme disturbances, the order fulfillment rate decrease is controlled within 7%, more than 50% lower than traditional solutions, significantly enhancing its robustness against disturbances.

[0074] This invention employs a closed-loop iterative optimization mechanism to maintain long-term prediction and scheduling efficiency. It constructs a closed-loop system of "prediction-scheduling-feedback-optimization" through an online learning and feedback module. This system collects prediction errors, scheduling performance deviations (timeliness, cost, and fulfillment rate deviations) and the latest logistics data in real time. It dynamically updates the weight parameters of the hierarchical spatiotemporal attention network using a small-batch gradient descent method, adjusting the feature weights of node state vectors, the coefficients of the association matrix calculation, and the knowledge graph update mechanism (time window, trigger threshold). The optimization process uses an incremental learning mode, ensuring uninterrupted platform operation and supporting parameter rollback. This solves the problems of traditional technologies lacking dynamic adjustment mechanisms and experiencing long-term decline in accuracy and efficiency. Actual verification shows that after 30 days of continuous operation, the prediction accuracy (MAPE) remains around 7%, an improvement of over 30% compared to solutions without closed-loop optimization, demonstrating long-term stable and efficient prediction and scheduling performance.

[0075] With high scalability and real-time response capabilities, this invention is adapted to large-scale logistics networks. Utilizing HBase distributed storage technology, it supports the storage and real-time updating of dynamic logistics knowledge graphs with tens of millions of nodes and edges. The time window can be flexibly configured from 15 minutes to 1 hour, and the update latency triggered by critical events is ≤10 seconds, adapting to the ever-expanding scale of logistics networks. The hierarchical spatiotemporal attention network training employs an early stopping mechanism to avoid overfitting. The discrete event simulation model uses a 1-minute time step to ensure simulation accuracy. Simultaneously, GPU clusters accelerate model computation and simulation derivation, with a single scenario simulation taking ≤30 minutes, meeting the real-time prediction and scheduling requirements of large-scale logistics networks and possessing strong engineering applicability.

[0076] The integrated design lowers the barrier to entry and improves operational efficiency. This platform integrates data processing, spatiotemporal prediction, simulation, and closed-loop optimization into one system. Each module is sequentially connected to form a complete data flow and optimization chain, eliminating the need to build multiple independent systems and reducing the complexity and cost of technology implementation. The generated flexible scheduling plan clearly includes resource allocation schemes, route planning results, and emergency response procedures, directly guiding actual logistics operations. Simultaneously, the platform supports users in expanding disturbance scenarios and scheduling rules to adapt to the personalized needs of different industries (e-commerce logistics, instant delivery, trunk transportation, etc.), effectively reducing scheduling costs (15%-20% lower than traditional solutions), shortening delivery time deviations (controlled within 5 minutes), and significantly improving the intelligent operation level and market competitiveness of the logistics network. Attached Figure Description

[0077] Figure 1 This is a diagram of the platform's overall architecture.

[0078] Figure 2 Flowchart of the dynamic feature perception module;

[0079] Figure 3This is a schematic diagram of the spatiotemporal prediction engine network structure;

[0080] Figure 4 This is a schematic diagram of the simulation and optimization process. Detailed Implementation

[0081] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the invention, and not all of them. Unless otherwise specified, the embodiments and features described in this application can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0082] It should be noted that if the embodiments of the invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0083] Furthermore, "multiple" refers to two or more. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of a person skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the invention.

[0084] Please see Figure 1-4

[0085] I. Implementation Environment Configuration

[0086] (I) Hardware Environment

[0087] This embodiment uses a distributed server cluster deployment, with the specific configuration as follows:

[0088] Data processing nodes: 8 servers with Intel Xeon Gold 6330 CPUs (2.0GHz, 24 cores), 128GB of memory, and 4TB SSDs, used for multi-source data fusion and dynamic knowledge graph updates;

[0089] Model computing nodes: 4 servers with NVIDIA A100 GPUs (40GB VRAM) are used for training and inference of the hierarchical spatiotemporal attention network and Monte Carlo simulation.

[0090] Storage nodes: 3 servers deploying the HBase distributed storage system, with the number of replicas set to 3, supporting real-time storage and high-concurrency access of tens of millions of logistics knowledge graph nodes and edges;

[0091] Network environment: The cluster is interconnected using 10Gbps Ethernet to ensure data transmission latency ≤5ms, meeting real-time processing requirements.

[0092] (II) Software Environment

[0093] Operating system: Ubuntu 20.04LTS Server 64-bit;

[0094] Data processing framework: Hadoop 3.3.4 (distributed data storage and computing), Spark 3.4.1 (multi-source data cleaning and transformation);

[0095] Deep learning framework: TensorFlow 2.10.0 (construction and training of hierarchical spatiotemporal attention networks);

[0096] Simulation tools: Python 3.9.16, SimPy 4.0.1 (for discrete event simulation models);

[0097] Optimization algorithm library: DEAP 1.3.3 (NSGA-III multi-objective optimization algorithm implementation);

[0098] Databases: HBase 2.4.14 (dynamic logistics knowledge graph storage), Redis 6.2.10 (real-time data cache).

[0099] II. Data Preparation and Preprocessing

[0100] (I) Multi-source data access

[0101] This embodiment integrates the following six types of heterogeneous logistics time-series data, covering key information across the entire logistics chain:

[0102] 1. Order data: Sourced from e-commerce platforms and instant logistics systems, including fields such as order number, receiving / sending node ID, cargo weight, timeliness requirements, and order timestamp. Data is collected at a rate of 1 record per second.

[0103] 2. GPS trajectory data: from the transport vehicle terminal, including vehicle ID, latitude and longitude coordinates, driving speed, real-time location timestamp, etc., collected at a frequency of 10 seconds / data.

[0104] 3. Warehouse inventory data: sourced from the Warehouse Management System (WMS), including warehouse ID, product SKU, inventory quantity, inbound time, outbound time, etc., collected every minute.

[0105] 4. Road network status data: sourced from the open API of the city traffic management department, including road segment ID, traffic status (smooth / slow / congested), average traffic speed, congestion index, etc., collected every 5 minutes.

[0106] 5. Weather data: Sourced from the meteorological service platform API, including region ID, temperature, precipitation probability, wind force level, extreme weather warning indicators, etc., collected every 15 minutes.

[0107] 6. Sales campaign data: Sourced from the e-commerce platform's marketing system, including campaign ID, coverage area, campaign start and end time, expected order increase percentage, etc., synchronized in real time according to the campaign cycle.

[0108] (II) Data Preprocessing Flow

[0109] Step 1, Data Cleaning: Abnormal data is handled using the following rules:

[0110] Remove invalid records from the order data that have "empty sender / receiver node ID" or "goods weight ≤ 0";

[0111] The 3σ criterion is used to identify speed anomalies in GPS trajectory data (those exceeding the historical speed mean of the road segment by ±3 times the standard deviation), and linear interpolation is used to fill in the missing values.

[0112] For abnormal records in the warehouse inventory data where the "inventory quantity is negative", corrections are made in conjunction with the inbound and outbound transaction records;

[0113] Step 2, Data Standardization: Using the Min-Max standardization method, all numerical features are mapped to the interval [0, 1]. The formula is as follows:

[0114]

[0115] Where x is the original feature value, x min x max These are the historical minimum and maximum values ​​of this feature, respectively.

[0116] Step 3, Spatiotemporal Alignment:

[0117] Time alignment: The timestamps of all data are unified in "seconds", and low-frequency data (such as weather and road network status) are padded to the second level using forward padding.

[0118] Spatial alignment: Based on the WGS-84 geocoding standard, the addresses of all nodes (warehouses, distribution centers, and sites) are converted into latitude and longitude coordinates (accurate to 6 decimal places) to ensure that the spatial location matching accuracy is ≤10 meters.

[0119] III. Specific Implementation Steps for Each Module

[0120] (I) Implementation of the Dynamic Feature Perception Module

[0121] Construction of dynamic logistics knowledge graph

[0122] Entity definition and attribute assignment:

[0123] Logistics node entities include warehouses (100), distribution centers (20), and stations (500). Attributes include node ID, latitude and longitude coordinates, inventory capacity (warehouse / distribution center), hourly processing capacity (sorting / outbound efficiency), current inventory level, and number of personnel.

[0124] Transportation route entities: a total of 1200 routes, with attributes including route ID, origin node ID, destination node ID, actual distance (unit: kilometers), designed transportation time (unit: minutes), and current capacity saturation (occupied capacity / total capacity);

[0125] Order Entity: Valid orders that are accessed in real time, with attributes including order ID, shipping node ID, receiving node ID, cargo weight, delivery time requirement, and order placement time;

[0126] Edge definition and association strength calculation:

[0127] "Node-Route" Edge: Connects logistics nodes and transportation routes. The edge attribute is the association strength, determined by the formula... Calculate, for example, the transportation distance d between node A (warehouse) and node B (distribution center). AB = 50 kilometers, transit time T AB =60 minutes, capacity saturation S AB =0.7, normalized value of historical collaboration frequency F AB =0.8, current order flow percentage O AB =0.6, then M AB =0.3 / 50+0.25 / 60+0.2*(1-0.7)+0.15*0.8+0.1*0.6≈0.327;

[0128] "Order-Node" edge: connects the order to the shipping / receiving node. The edge attribute is the order's contribution to the node's demand (calculated as a percentage of the order's weight).

[0129] Knowledge graph update mechanism execution

[0130] Sliding update with time window: The default 30-minute time window is used. Every 30 minutes, all entity attributes (such as node inventory level and line capacity saturation) and edge association strength in the knowledge graph are recalculated and updated.

[0131] Critical event-triggered update: Set trigger thresholds: ① Order volume exceeds twice the historical average for the same period (historical average is based on statistics of the same period in the last 30 days); ② Weather platform issues warnings for extreme weather such as heavy rain or blizzard; ③ Node processing capacity decreases by more than 30% (e.g., equipment failure). When any of the thresholds are met, a full map update is immediately triggered with an update delay of ≤10 seconds.

[0132] Spatiotemporal dynamic feature extraction

[0133] Node state vector: Extract 12-dimensional features, with dimensions of 530×12 (530 logistics nodes, 12-dimensional features). The features of each dimension include: node inventory level (after standardization), processing capacity utilization rate, number of orders in transit, average demand in the past 24 hours, weather impact coefficient (0-1, 1 for extreme weather), road network congestion index, staffing adequacy rate, equipment operating status, promotional activity impact coefficient, historical demand volatility, node priority, and emergency reserve level.

[0134] Association matrix: 530×530, matrix elements are the association strength M between corresponding nodes. ij The association strength of nodes without direct transport routes is set to 0.

[0135] (II) Implementation of the Spatiotemporal Prediction Engine

[0136] Training of hierarchical spatiotemporal attention networks

[0137] Training dataset: Preprocessed time series data from the past 90 days is used, divided into training, validation, and test sets in a 7:2:1 ratio. The training batch size is set to 64, the maximum number of iterations is 100, and an early stopping mechanism is adopted (training stops if the validation set error does not decrease for 5 consecutive iterations).

[0138] Network parameter initialization: weight matrix W h (12x64), W a (128×1) The Xavier uniform initialization is used, the attention coefficient vector a is initialized to a 1×1 random value, the activation function is ReLU, and the dropout rate is set to 0.2 to prevent overfitting.

[0139] Spatial attention subnetwork computation

[0140] Step S1: For the node state vector h i (12-dimensional) linear transformation, h i ′=W h ·h i Output a 64-dimensional node feature vector;

[0141] Step S2: Calculate the attention score e between nodes i and j. ij =LeakyReLU(W a [h i ′||h j ′]·a), where [h i ′||h j ′] is the concatenation of the feature vectors of nodes i and j (128 dimensions), and the negative slope of LeakyReLU is set to 0.01;

[0142] Step S3: Based on the correlation matrix M ij Filtering invalid associations (M) ij =0 e ij =0), and the attention weight α is obtained by softmax normalization. ij =softmax j (e ij Ensure that the sum of the weights is 1;

[0143] Step S4: Weighted aggregation of neighbor node features Generate a 64-dimensional spatial context feature vector.

[0144] Temporal attention subnetwork computation

[0145] Input historical hidden state sequence: Take the spatial feature vector sequence of the last 24 hours (T=24).

[0146] Step S1: Calculate the attention weights at time t' using a 2-layer perceptron (MLP). The first layer of the MLP outputs 32-dimensional features, and the second layer outputs 1-dimensional weight values.

[0147] Step S2: Temporal Feature Aggregation A 64-dimensional temporal feature vector is obtained.

[0148] Predicted output

[0149] Feature fusion: concatenating spatial features with temporal features. (128 dimensions);

[0150] Decoding Output: Input a fully connected layer (128×12) and a temporal GRU decoding layer (64 hidden layer dimensions, 12 iteration steps), output the demand prediction sequence y for each node in the next 12 hours (K=12). i (t+1), ..., y i (t+12), the forecast result is the standardized demand value, which is converted into the actual demand quantity (such as order quantity, cargo weight) through inverse standardization.

[0151] (III) Implementation of Simulation and Optimization Module

[0152] Construction of Discrete Event Simulation Model for Logistics Network

[0153] Implemented using the SimPy framework, the model's topology fully maps to the dynamic logistics knowledge graph, and includes the following core processes:

[0154] Node processing flow: Inbound (processing time follows a log-normal distribution LN(μ=3, σ2 =0.5))→ Sorting (processing time follows a log-normal distribution LN(μ=2,σ 2 =0.3))→Outbound (processing time follows a log-normal distribution LN(μ=1.5,σ 2 =0.2);

[0155] Transportation process: Loading (fixed time 5 minutes) → On the way (time = route distance / real-time driving speed, real-time driving speed is dynamically adjusted based on road network conditions) → Unloading (fixed time 3 minutes);

[0156] Resource constraints: The number of vehicles is configured according to node level (warehouse ≥ 50 vehicles, distribution center ≥ 20 vehicles, station ≥ 5 vehicles). Warehouse capacity and personnel configuration are set as hard constraints according to entity attributes. Resource scheduling warnings are triggered when the constraints are exceeded.

[0157] Basic scheduling plan generation

[0158] By combining five predefined scheduling rules, five basic contingency plans are generated:

[0159] Contingency Plan 1 (Cost Priority): Use an enumeration method to select the resource allocation scheme with the lowest transportation cost (based on distance × unit mileage freight) + warehousing cost (inventory occupancy fee × storage time), and prioritize routes with low capacity saturation;

[0160] Contingency Plan 2 (Time-sensitive): Calculate the shortest delivery route based on Dijkstra's algorithm, prioritize the use of high-speed vehicles, and allow for a moderate increase in costs to meet time-sensitive requirements;

[0161] Contingency Plan 3 (Balanced Type): Construct an objective function min(0.5 x capacity utilization deviation + 0.5 x timeliness deviation) and solve for the balanced solution;

[0162] Contingency Plan 4 (Emergency Priority): Reserve 10% of vehicle and warehouse space, and adopt a nearby allocation strategy for sudden needs;

[0163] Contingency Plan 5 (Regional Collaboration): Divide all nodes into 8 regions, with nodes within each region sharing 30% of the transportation capacity. Cross-regional transportation will prioritize the regional transit mode.

[0164] External disturbance scene injection and Monte Carlo simulation

[0165] Disturbance scenario parameter configuration:

[0166] Extreme weather: The traffic efficiency of the core route (accounting for 20% of the total route) will randomly decrease by 30%-50%, and the duration will be randomly selected as 2-4 hours;

[0167] Node failure: Randomly select 1-2 distribution centers, and the processing capacity will randomly decrease by 50%-100% for a random duration of 1-3 hours;

[0168] Demand surge: Randomly select a region, and the demand of local nodes will surge by 100%-200% for a random duration of 1-2 hours;

[0169] Disturbance Combination and Simulation Execution: Each basic contingency plan is injected with 27 random combinations of 3 scenarios. Each combination is executed 2000 Monte Carlo simulations. The delivery time deviation, scheduling cost, and order fulfillment rate data of each simulation are recorded. The simulation time step is set to 1 minute, and the simulation time for a single contingency plan is ≤30 minutes.

[0170] Multi-objective optimization and plan ranking

[0171] Performance index calculation:

[0172] N is the total number of orders;

[0173] Dispatch cost C = Transportation cost + Warehousing cost + Emergency cost (emergency cost is calculated based on emergency resource occupancy fee);

[0174] Constraint R ≥ 90%;

[0175] Multi-objective optimization: Using the NSGA-III algorithm, with a population size of 200, 100 iterations, a crossover probability of 0.8, and a mutation probability of 0.1, the Pareto optimal solution set is obtained;

[0176] Contingency plan ranking: Calculate the comprehensive score based on "timeliness deviation weight 0.4 + cost weight 0.3 + satisfaction rate weight 0.3", select the top 3 flexible scheduling contingency plans from the Pareto optimal solution set, and output the contingency plan details (including resource allocation scheme, path planning results, and emergency response process).

[0177] (iv) Implementation of the online learning and feedback module

[0178] Data acquisition and monitoring

[0179] Real-time collection of three types of data:

[0180] Prediction error This represents the actual required value for node i.

[0181] Execution effect deviation ΔE=|E 实际 -E 仿真 |、ΔC=|C 实际 -C 仿真 |、ΔR=|R 实际 -R 仿真 |;

[0182] Latest logistics data: preprocessed time-series data including new orders, real-time road network status, and changes in node inventory;

[0183] Data storage: Collected data is partitioned and stored in HBase according to timestamps, and 90 days of historical data are retained for model optimization.

[0184] Adaptive optimization execution

[0185] Model parameter optimization: Based on prediction error and new data, mini-batch gradient descent (batch size 32, learning rate 0.001) is used to update the weight parameters (W) of the hierarchical spatiotemporal attention network. h W a a) Dynamically adjust the historical time series length T (range 12-48) and the prediction step size K (range 6-24). For example, when the average prediction error in the past 24 hours is ≥12%, adjust T to 36 and K to 18.

[0186] Feature extraction rule optimization: When ΔE ≥ 10 minutes, increase the weight of the weather influence coefficient in the node state vector from 0.1 to 0.15; when ΔC ≥ 5%, increase the weight of the transportation distance in the association matrix from 0.3 to 0.35.

[0187] Knowledge graph update mechanism optimization: When the frequency of key events is too high (≥5 times per day), the threshold for sudden increase in order volume will be adjusted from 2 times the historical average to 1.8 times; when the update delay is ≥15 seconds, the time window will be shortened from 30 minutes to 20 minutes.

[0188] Iteration mechanism: The default optimization cycle is 24 hours. When the average prediction error exceeds 15%, immediate optimization is triggered. The optimization process adopts incremental learning, only updating some parameters of the model without interrupting the platform operation. The optimization time is ≤10 minutes. At the same time, the parameter version of each optimization is saved, and it is supported to roll back to the most recent optimal version when the optimization effect is not good.

[0189] IV. Implementation Results Verification

[0190] (I) Verification Indicators and Comparison Objects

[0191] Key validation metrics: Prediction accuracy (Mean Absolute Percentage Error, MAPE), delivery timeliness deviation, scheduling cost, order fulfillment rate, and disturbance resistance (the extent to which the fulfillment rate decreases under disturbance scenarios).

[0192] Comparison objects: ① Traditional prediction models (ARIMA, LSTM); ② Existing static scheduling schemes; ③ Simplified platform without online learning mechanism.

[0193] (II) Verification Results

[0194] Comparison of prediction accuracy:

[0195] Platform of this invention: MAPE = 8.3%;

[0196] LSTM model: MAPE = 15.7%;

[0197] ARIMA model: MAPE = 21.4%;

[0198] It is evident that the hierarchical spatiotemporal attention network of this invention has a superior ability to capture spatiotemporal dependencies and significantly improves prediction accuracy.

[0199] Scheduling performance comparison (normal scenario):

[0200]

[0201] Comparison of disturbance resistance capabilities (extreme weather + sudden demand changes):

[0202] The platform of this invention: order fulfillment rate decreased to 92.3% (a decrease of 6.8%), and delivery time deviation increased to 7.5 minutes;

[0203] Static scheduling scheme: Order fulfillment rate dropped to 82.1% (a decrease of 13.2%), and delivery time deviation increased to 15.3 minutes;

[0204] Verification shows that the flexible scheduling scheme of the present invention can effectively resist external disturbances and has stronger robustness.

[0205] Long-term iteration effect:

[0206] After 30 days of continuous operation, the MAPE of the platform of this invention decreased from the initial 8.3% to 7.1% due to the continuous optimization of the online learning mechanism, while the MAPE of the simplified platform without online learning increased to 10.5%, proving that the closed-loop iterative mechanism can maintain long-term prediction accuracy and scheduling efficiency.

[0207] In summary, this invention effectively solves the problems of insufficient data fusion, inaccurate capture of spatiotemporal dependencies, and weak anti-disturbance capability in traditional logistics demand forecasting and scheduling through multi-module collaboration and closed-loop optimization, and significantly improves the intelligent operation level of logistics networks.

[0208] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0209] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0210] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A big data-driven logistics demand forecasting platform, characterized in that, include: The dynamic feature perception module is used to access and fuse multi-source heterogeneous logistics time-series data, construct and update a dynamic network model that represents the state and relationship of the logistics network in real time, and extract spatiotemporal dynamic features from the dynamic network model. The spatiotemporal prediction engine, connected to the dynamic feature perception module, is used to receive the spatiotemporal dynamic features and output the demand prediction sequence of each logistics node in multiple future time periods by fusing the spatiotemporal dependence of the features. The simulation and optimization module, connected to the spatiotemporal prediction engine, is used to receive the demand prediction sequence and construct a logistics network simulation model based on the current network state output by the dynamic feature perception module. By simulating various preset external disturbance scenarios in the simulation model, the performance of different scheduling strategies under uncertainty is evaluated, and a set of elastic scheduling plans sorted by priority is output.

2. The big data-driven logistics demand forecasting platform according to claim 1, characterized in that: The dynamic network model is a dynamic logistics knowledge graph; the dynamic feature perception module updates the structure and attributes of the dynamic logistics knowledge graph based on a preset time window sliding or key event triggering mechanism. The extracted spatiotemporal dynamic features include at least: node state vectors representing the state of each logistics node at time t, and correlation matrices representing the correlation strength between nodes at time t.

3. The big data-driven logistics demand forecasting platform according to claim 1, characterized in that: The spatiotemporal prediction engine incorporates an attention-based prediction model, which includes: The spatial attention unit is used to calculate the dynamic spatial dependencies between nodes and generate hidden node states containing spatial context information, taking the node state vector and the correlation matrix as input. The temporal attention unit is used to adaptively capture the temporal patterns of state evolution by taking the historical hidden state sequence of each node as input. The prediction output unit is used to fuse the features processed by the spatial attention unit and the temporal attention unit, and decode and output the demand prediction sequence.

4. The big data-driven logistics demand forecasting platform according to claim 1, characterized in that: The prediction model based on the attention mechanism is a hierarchical spatiotemporal attention network, the spatial attention unit is a graph attention subnetwork, and the temporal attention unit is a temporal attention subnetwork.

5. The big data-driven logistics demand forecasting platform according to claim 1, characterized in that, The simulation and optimization module is specifically used for: Construct a discrete event simulation model for a logistics network; Using the demand forecast sequence as the basic input and combining it with predefined scheduling rules, a variety of basic scheduling plans are generated. In the discrete event simulation model, various external disturbance scenarios, including extreme weather, node failure, or sudden demand changes, are injected into each basic scheduling plan, and Monte Carlo simulation is performed. Based on multiple performance indicators from simulation output, a multi-objective optimization method is used to comprehensively evaluate and rank the plans, generating the set of flexible scheduling plans.

6. The big data-driven logistics demand forecasting platform according to any one of claims 1-5, characterized in that, The platform also includes: The online learning and feedback module is used to continuously collect the actual prediction error of the spatiotemporal prediction engine, the execution effect deviation of the elastic scheduling plan set, and the latest logistics data. It uses this data to adaptively optimize the model parameters of the spatiotemporal prediction engine and the feature extraction rules of the dynamic feature perception module, thereby realizing the closed-loop iteration of the platform.

Images