Scene demand peak valley prediction method and system based on large space-time model

Abstract

The invention provides a scene demand peak-valley prediction method and system based on a space-time large model, and relates to the technical field of intelligent traffic, and the method comprises the steps: obtaining multi-source heterogeneous space-time data of a target region, carrying out the non-uniform space grid division, calculating an instantaneous supply-demand ratio and a settlement price rolling fluctuation ratio, and constructing a settlement risk enhanced space-time feature tensor; inputting the settlement risk enhanced spatio-temporal feature tensor into a peak-valley prediction spatio-temporal large model, extracting features through a spatio-temporal diagram attention network fusing settlement price fluctuation features, and respectively outputting a transportation order demand prediction value and an uncertainty quantitative index by a double-branch decoder; inputting the transportation order demand predicted value and the uncertainty quantitative index into a strategy to generate a cellular automaton, simulating supply and demand propagation by dynamically adjusting state transfer resistance between cells, and generating a non-uniform distributed settlement parameter adjustment instruction; and directly outputting a settlement parameter configuration result of each sub-region in the target region based on the distributed settlement parameter adjustment instruction.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically to a method and system for predicting peak and valley demand based on a spatiotemporal large model. Background Technology

[0002] In the field of intelligent transportation, accurately predicting the peak and trough of regional transportation demand and dynamically adjusting settlement parameters is a core link in balancing supply and demand and stabilizing the market. It is also a key application scenario for the deep integration of artificial intelligence, big data, and Internet of Things technologies in the new generation of information technology industry with intelligent manufacturing equipment and intelligent scheduling systems in the high-end equipment manufacturing industry.

[0003] However, existing technologies have obvious limitations.

[0004] First, traditional spatial grid division methods are mostly uniform or static, failing to combine the heterogeneity of urban functional area distribution with the dynamic characteristics of spatiotemporal big data. They cannot accurately reflect the real supply and demand heterogeneity caused by the uneven distribution of urban functional areas, resulting in biases in the prediction basis.

[0005] Second, mainstream forecasting models largely rely on basic characteristics such as historical order volume, failing to effectively integrate the key risk indicator of "settlement price fluctuations," resulting in insufficient robustness in forecasting during periods of sharp price volatility. Furthermore, these models typically output only a single forecast value, lacking quantification of forecast uncertainty, thus limiting the risk flexibility of downstream strategy formulation.

[0006] Third, in terms of strategy generation, methods based on centralized optimization or static thresholds are difficult to simulate the dynamic propagation process of supply and demand in a non-uniform space, and cannot generate distributed settlement adjustment instructions that accurately match the unique situation of each sub-region.

[0007] Therefore, we now provide a method and system for predicting peak and valley demand based on a spatiotemporal large model, which can achieve refined, risk-resistant, and quantifiable demand forecasting and dynamic settlement control, greatly improving the scientific nature of forecasting and control. Summary of the Invention

[0008] To address the aforementioned technical problems, the present invention aims to provide a method and system for predicting peak and valley demand based on a spatiotemporal large model. This method enables refined, risk-resistant, and quantifiable demand forecasting and dynamic settlement control, thereby greatly improving the scientific nature of forecasting and control.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting peak and valley demand based on a spatiotemporal large model, comprising: Acquire multi-source heterogeneous spatiotemporal data of the target region, based on weighted... The graph algorithm is used to divide the space into non-uniform grids, calculate the instantaneous supply-demand ratio and the rolling volatility of the settlement price, and construct a settlement risk-enhanced spatiotemporal feature tensor. The settlement risk-enhanced spatiotemporal feature tensor is input into the pre-trained peak-valley prediction spatiotemporal model. Features are extracted by a spatiotemporal graph attention network that integrates settlement price fluctuation features. The dual-branch decoder outputs the predicted value of transportation order demand and the corresponding uncertainty quantification index, respectively. The predicted demand value of the transportation orders and the uncertainty quantification index are input into the strategy to generate a cellular automaton. By dynamically adjusting the state transmission resistance between cells, the supply and demand propagation is simulated, and non-uniform distributed settlement parameter adjustment instructions are generated. Based on the distributed settlement parameter adjustment instruction, the settlement parameter configuration results of each sub-region within the target area are directly output.

[0010] Preferably, the multi-source heterogeneous spatiotemporal data includes historical transportation order data, historical settlement price data, and real-time traffic status data.

[0011] Preferably, the weighted average The graph algorithm performs non-uniform spatial grid partitioning, calculates the instantaneous supply-demand ratio and settlement price rolling volatility, and constructs a settlement risk-enhanced spatiotemporal feature tensor, including: Based on the historical transportation order data and historical settlement price data, the corresponding historical order density distribution and historical settlement price volatility distribution are calculated respectively. Through weighting The graph algorithm divides the target region into a non-uniform spatial grid to generate various dynamic spatial units. Centered on the actual settlement time of historical transportation orders, a preset time window is traced back to extract the multi-source heterogeneous spatiotemporal data, forming a time-series data slice based on the settlement event; Within each of the dynamic spatial units, for each of the time-series data slices, higher-order features, including at least the instantaneous supply-demand ratio and the rolling volatility of the settlement price, are calculated; The dynamic spatial units, time-series data slices, and higher-order features are integrated to form a corresponding settlement risk-enhanced spatiotemporal feature tensor.

[0012] Preferably, the training process of the pre-trained spatiotemporal model for peak-valley prediction includes: Based on a pre-set attention mechanism library and decoder structure library, the attention mechanism in the attention mechanism library is paired with the decoder in the decoder structure library to form basic combination models of different categories. Each basic combination model contains a spatiotemporal feature decoupling module, which is used to separate static spatial features and dynamic temporal features through a gating mechanism. Based on the risk sensitivity coefficient, N different hyperparameter configurations are dynamically generated for each basic combination model; Retrieve multi-source heterogeneous spatiotemporal historical data corresponding to the target region in the database, preprocess the multi-source heterogeneous spatiotemporal historical data to obtain the corresponding settlement risk enhanced spatiotemporal feature historical tensor, and label the settlement risk enhanced spatiotemporal feature historical tensor. Based on the volatility of historical settlement prices, the business risk value of each time segment in the sequence of the historical tensor of the settlement risk-enhanced spatiotemporal characteristics is calculated, and all historical time segments are divided into multiple risk levels according to the distribution of business risk values. From each risk level, time segments are randomly selected according to a preset ratio to form a risk panorama verification set; the remaining time segments form a risk course training set. The time segments in the risk course training set are sorted from low to high according to their business risk values ​​to construct a training course with increasing risk. According to the training course order of increasing risk, the sorted time segments are sequentially input into the basic combination model for training, and the trained basic combination models are output; wherein, the spatiotemporal feature decoupling module dynamically adjusts the fusion weight of static spatial features and dynamic temporal features in its gating mechanism according to the business risk value of the currently input time segment. The trained basic combination models are evaluated on the risk panorama validation set, and the model with the best performance is selected as the pre-trained spatiotemporal large model for peak and valley prediction.

[0013] Preferably, the step of sequentially inputting the sorted time segments into the basic combination model for training according to the training course order of increasing risk, and outputting the trained basic combination models, includes: Based on the training course sequence with increasing risk, the sorted time segments are divided into multiple consecutive course stages, each course stage containing a set of time segments with increasing risk values; For each time segment in each course stage, the corresponding business risk value is read, and the business risk value is concatenated with the embedding code of the current course stage. The result is input into the gated neural network, and the output is the fusion weight. Based on the fusion weight, the static spatial features and dynamic temporal features are weighted and fused to form the fusion features. The fused features are input into the encoder and the dual-branch decoder to obtain the demand forecast and the uncertainty quantification index, respectively. Calculate the loss of the demand forecasting branch and the loss of the uncertainty quantification branch, and dynamically adjust the loss weights according to the current course stage to obtain the total loss. The gradient is calculated based on the total loss, and the model parameters in the basic combined model are updated. Only when the prediction error and uncertainty calibration reach the stage advancement threshold at the same time will the model enter the next course stage. After all course stages are completed, the corresponding basic combined model will be the trained basic combined model.

[0014] Preferably, selecting the model with the best performance as the pre-trained spatiotemporal large model for peak-valley prediction includes: A three-dimensional evaluation space is constructed, the dimensions of which include prediction accuracy dimension, risk robustness dimension, and strategy value dimension; Based on the aforementioned risk panorama validation set, all trained basic combinatorial models are evaluated in the three-dimensional evaluation space to identify all non-dominated solutions and form a Pareto front. Calculate the proportion of high-risk samples in the risk panorama verification set and use it as the corresponding risk exposure coefficient; Based on the risk exposure coefficient, the coordinates of the reference point are dynamically determined in the three-dimensional assessment space: Calculate the weighted Chebyshev distance from each fundamental composite model on the Pareto front to the reference point, and select the fundamental composite model with the smallest weighted Chebyshev distance as the model with the best performance.

[0015] Preferably, the step of inputting the predicted demand value of the transportation order and the uncertainty quantification index into the strategy to generate a cellular automaton, and simulating supply and demand propagation by dynamically adjusting the state transmission resistance between cells to generate non-uniform distributed settlement parameter adjustment instructions includes: The target area is divided into multiple interconnected cellular grids, each cellular grid representing a geographic sub-region, and a state vector is maintained, which includes the current supply-demand ratio, historical settlement price, driver density, and passenger waiting time. Based on the predicted demand for transportation orders, each cell grid interacts with its neighboring cell grids according to preset neighbor interaction rules to simulate the dynamic propagation process of supply and demand. The resistance coefficient for state transfer between cell grids is dynamically adjusted using uncertainty quantification indicators. After multiple iterations, the stable states of each cell grid are collected, and a non-uniform distributed settlement parameter adjustment strategy is generated based on the principle of spatial consistency. In this way, each sub-region within different target regions obtains differentiated distributed settlement parameter adjustment instructions.

[0016] A second aspect of the present invention also provides a scenario demand peak and valley prediction system based on a spatiotemporal large model, comprising: The acquisition and preprocessing module is used to acquire multi-source heterogeneous spatiotemporal data of the target area, based on weighted... The graph algorithm is used to divide the space into non-uniform grids, calculate the instantaneous supply-demand ratio and the rolling volatility of the settlement price, and construct a settlement risk-enhanced spatiotemporal feature tensor. The peak-valley prediction module is used to input the settlement risk-enhanced spatiotemporal feature tensor into the pre-trained peak-valley prediction spatiotemporal model, extract features through a spatiotemporal graph attention network that integrates settlement price fluctuation features, and output the predicted value of transportation order demand and the corresponding uncertainty quantification index by a dual-branch decoder. The adjustment instruction module is used to input the predicted value of the transportation order demand and the uncertainty quantification index into the strategy to generate a cellular automaton. By dynamically adjusting the state transmission resistance between cells to simulate supply and demand propagation, it generates non-uniform distributed settlement parameter adjustment instructions. The calculation parameter output module is used to directly output the settlement parameter configuration results of each sub-region within the target area based on the distributed settlement parameter adjustment instruction.

[0017] Compared with existing technologies, the beneficial effects of this invention are: the "Scenario Demand Peak and Valley Prediction Method Based on Spatiotemporal Large Model" provided by this invention constructs a complete technical closed loop from refined data perception to intelligent risk prediction, and then to precise adaptive regulation. Compared with existing technologies, its overall solution brings revolutionary progress: Through weighting The graph algorithm divides the target area into a non-uniform spatial grid and innovatively calculates the instantaneous supply-demand ratio and the rolling volatility of settlement prices, constructing a spatiotemporal feature tensor enhanced with settlement risk. The non-uniform division ensures that the dynamic spatial units closely align with the actual functional areas and business density of the city, eliminating data bias caused by unreasonable grid division at its source. By using the rolling volatility of settlement prices, a key financial risk indicator, as a core feature, the subsequent model can "see" market instability during the data input stage, providing a fundamental guarantee for making robust predictions.

[0018] A risk-based training strategy is employed, utilizing a dual-branch decoder to simultaneously output demand forecasts and uncertainty quantification indicators. Finally, the final model is selected based on the Pareto frontier in a three-dimensional evaluation space. Risk-based training allows the model to learn progressively, gradually overcoming high-volatility scenarios and significantly improving forecast stability and generalization ability during critical peak and trough periods. The dual-branch decoder provides a complete information package of "forecast value + uncertainty," providing a direct basis for generating resilient strategies. The three-dimensional evaluation model, which integrates forecast accuracy, risk robustness, and strategy value, ensures that the final selected model is not only theoretically accurate but also offers the highest overall value in business practice.

[0019] The prediction results are input into a cellular automaton to generate a strategy. By dynamically adjusting the resistance to state transmission between cells, supply and demand propagation is simulated, generating non-uniform distributed settlement parameter adjustment instructions. The cellular automaton simulates the propagation and interaction of supply and demand between adjacent regions, enabling the generated strategy to reflect real spatiotemporal diffusion effects and making the strategy more scientific. Utilizing uncertainty quantification indicators to dynamically adjust the resistance between cells allows the strategy system to actively intervene when predictions are reliable and automatically conserve when predictions are uncertain, preventing the spread of erroneous strategies and possessing strong risk resilience. The final generated distributed settlement parameter adjustment instructions can customize control parameters for each sub-region, achieving a fundamental improvement in operational scheduling from extensive to intensive, maximizing control efficiency and effectiveness. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1 This is a schematic diagram of the principle of a scenario demand peak and valley prediction method based on a spatiotemporal large model.

[0022] Figure 2 This is a schematic diagram of a scenario demand peak and valley prediction system based on a spatiotemporal large model. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] Example 1 like Figure 1 As shown, this embodiment discloses a method for predicting peak and valley demand in a scenario based on a spatiotemporal large model, including: Acquire multi-source heterogeneous spatiotemporal data of the target region, based on weighted... The graph algorithm is used to divide the space into non-uniform grids, calculate the instantaneous supply-demand ratio and the rolling volatility of the settlement price, and construct a settlement risk-enhanced spatiotemporal feature tensor. It should be noted that the multi-source heterogeneous spatiotemporal data includes historical transportation order data, historical settlement price data, and real-time traffic status data; The weighted The graph algorithm performs non-uniform spatial grid partitioning, calculates the instantaneous supply-demand ratio and settlement price rolling volatility, and constructs a settlement risk-enhanced spatiotemporal feature tensor, including: Based on the historical transportation order data and historical settlement price data, the corresponding historical order density distribution and historical settlement price volatility distribution are calculated respectively. Through weighting The graph algorithm divides the target region into a non-uniform spatial grid to generate various dynamic spatial units. Centered on the actual settlement time of historical transportation orders, a preset time window is traced back to extract the multi-source heterogeneous spatiotemporal data, forming a time-series data slice based on the settlement event; Within each dynamic spatial unit, for each time-series data slice, higher-order features are calculated, including at least the instantaneous supply-demand ratio and the rolling volatility of settlement prices. These higher-order features include the instantaneous supply-demand ratio (order demand / capacity supply within the current unit), the rolling volatility of settlement prices (the standard deviation of price fluctuations over the past N settlement periods), and the spatiotemporal correlation risk coefficient (order flow intensity × price volatility correlation between the current unit and surrounding units). In this embodiment, the higher-order features are also fused with basic spatiotemporal features, including time features (such as weekdays / weekends, peak hours / off-peak hours) and spatial features (such as the coordinates of the unit's center point and the surrounding traffic congestion index).

[0026] The dynamic spatial units, time-series data slices, and higher-order features are integrated to form a corresponding settlement risk-enhanced spatiotemporal feature tensor.

[0027] In this embodiment, the integration process is as follows: Construct a four-dimensional tensor, specifically as follows: ; T represents the number of time-series data slices generated through the settlement cycle alignment sub-step; N represents the total number of dynamic spatial units generated through the dynamic spatial gridding sub-step; C represents the total dimension of the feature channels; By traversing all T time slices and N dynamic spatial units, a C-dimensional feature vector f_tn is assigned to each spatiotemporal point (t, n). This feature vector is formed by concatenating the core business features, original observation features, and metadata embedding features at the corresponding spatiotemporal point.

[0028] The settlement risk-enhanced spatiotemporal feature tensor is input into the pre-trained peak-valley prediction spatiotemporal model. Features are extracted by a spatiotemporal graph attention network that integrates settlement price fluctuation features. The dual-branch decoder outputs the predicted value of transportation order demand and the corresponding uncertainty quantification index, respectively. It should be noted that the training process of the pre-trained spatiotemporal model for peak-valley prediction includes: Based on a pre-set attention mechanism library and decoder structure library, the attention mechanisms in the attention mechanism library are paired with the decoders in the decoder structure library to form basic combined models of different categories. Each basic combined model includes a spatiotemporal feature decoupling module, which is used to separate static spatial features and dynamic temporal features through a gating mechanism. In this embodiment, the attention mechanism library includes three types of mechanisms: spatiotemporal graph attention (GAT), multi-head self-attention, and cross-attention. The decoder structure library includes three types of structures: fully connected decoder, LSTM decoder, and Transformer decoder.

[0029] Based on the risk sensitivity coefficient, N different hyperparameter configurations are dynamically generated for each basic combination model; Retrieve multi-source heterogeneous spatiotemporal historical data corresponding to the target region in the database, preprocess the multi-source heterogeneous spatiotemporal historical data to obtain the corresponding settlement risk enhanced spatiotemporal feature historical tensor, and label the settlement risk enhanced spatiotemporal feature historical tensor. Based on the volatility of historical settlement prices, the business risk value for each time segment in the sequence of the historical tensor of the settlement risk-enhanced spatiotemporal feature is calculated, and according to the distribution of business risk values, all historical time segments are divided into multiple risk levels. In this embodiment, the formula for calculating the business risk value is as follows: ; ; In the formula, This represents the original business risk value within the time segment corresponding to time t. This represents the normalized business risk value for the time segment corresponding to time t. Let be the rolling volatility of the settlement price within the time segment corresponding to time t, and let be the standard deviation of the settlement price within the rolling window corresponding to that time segment. is the supply-demand imbalance coefficient within the time segment corresponding to time t. It represents the degree of deviation between order demand and transportation capacity supply within that time segment; a larger value indicates a greater supply-demand imbalance and correspondingly higher business risk. Let be the spatiotemporal correlation risk coefficient within the time segment corresponding to time t. It is the product of the "order flow intensity between the current dynamic spatial unit and surrounding units" and the "price fluctuation correlation coefficient" within that time segment. This is a complete set of historical original business risk values. It consists of all historical time segments from the target region. (Original business risk value) consists of (in set form) (where m is the total number of historical time segments), is The benchmark dataset is normalized. The formula for calculating the rolling volatility of settlement prices within the time segment corresponding to time t is: ; In the formula, For a rolling window (e.g., 5 settlement cycles). For the first Periodic settlement price, This represents the average price within the window, reflecting the risk of price fluctuations.

[0030] The formula for calculating the spatiotemporal correlation risk coefficient within the time segment corresponding to time t is: ; This represents the order flow intensity between the current unit and surrounding units (e.g., the number of cross-unit orders per hour / total number of orders). This is the correlation coefficient between the price fluctuations of the current unit and its surrounding units (reflecting risk transmission).

[0031] From each risk level, time segments are randomly selected according to a preset ratio to form a risk panorama verification set; the remaining time segments form a risk course training set. The time segments in the risk course training set are sorted from low to high according to their business risk values ​​to construct a training course with increasing risk. According to the training course order of increasing risk, the sorted time segments are sequentially input into the basic combination model for training, and the trained basic combination models are output; wherein, the spatiotemporal feature decoupling module dynamically adjusts the fusion weight of static spatial features and dynamic temporal features in its gating mechanism according to the business risk value of the currently input time segment. It should be noted that, according to the training course order of increasing risk, the sorted time segments are sequentially input into the basic combination model for training, and the trained basic combination models are output, including: Based on the training course sequence with increasing risk, the sorted time segments are divided into multiple consecutive course stages, each course stage containing a set of time segments with increasing risk values; For each time segment in each course stage, the corresponding business risk value is read, and the business risk value is concatenated with the embedding code of the current course stage. The result is input into the gated neural network, which outputs fusion weights. Based on the fusion weights, the static spatial features and dynamic temporal features are weighted and fused to form fused features. In this embodiment, the gated neural network adopts a two-layer fully connected network (MLP) structure. The core is to output fusion weights to adjust the proportion of static and dynamic features.

[0032] The first layer uses ReLU as the activation function. The second layer uses Sigmoid as the activation function.

[0033] The fused features are input into the encoder and the dual-branch decoder to obtain the demand forecast and the uncertainty quantification index, respectively. Calculate the loss of the demand forecasting branch and the loss of the uncertainty quantification branch, and dynamically adjust the loss weights according to the current course stage to obtain the total loss. In this embodiment, the loss value corresponding to the demand forecasting branch is calculated using the mean absolute percentage error (MASE). The loss value corresponding to the uncertainty quantification branch is calculated using heteroscedastic regression loss.

[0034] The gradient is calculated based on the total loss, and the model parameters in the basic combined model are updated. Only when both the prediction error and uncertainty calibration reach the stage advancement threshold does the model proceed to the next stage. This process continues until all stage advancements are completed, at which point the corresponding basic combined model is considered fully trained. In this embodiment, the model parameters include gated neural network parameters, encoder parameters (attention weights and graph convolution kernel weights of the spatiotemporal graph attention network (GAT)), dual-branch decoder parameters (weights / biases of the fully connected layers in the demand prediction branch and uncertainty quantization branch), and feature mapping parameters (if the static / dynamic feature dimensions do not match).

[0035] The trained basic combination models are evaluated on the risk panorama validation set, and the model with the best performance is selected as the pre-trained spatiotemporal large model for peak and valley prediction.

[0036] It should be noted that the selection of the best-performing model as the pre-trained spatiotemporal large-scale peak-valley prediction model includes: A three-dimensional evaluation space is constructed, comprising prediction accuracy, risk robustness, and strategy value dimensions. In this embodiment, the prediction accuracy dimension focuses on the model's prediction error of transportation order demand, using the reciprocal of the Mean Absolute Percentage Error (MAPE) as a quantification indicator. The specific calculation formula is as follows: ; In the formula, For the dimension of prediction accuracy, where, The calculation needs to distinguish between risk levels and weighting (to avoid low-risk samples dominating the error).

[0037] The formula for calculating the strategy value dimension is: ; In the formula, To avoid using model prediction, historical settlement parameters (such as settlement prices in the same region last week) are used to calculate the corresponding business revenue. This revenue = order completion volume × unit settlement profit - empty running cost; Based on the demand forecast values ​​(adjusting settlement prices to match demand peaks and troughs) and uncertainty quantification indicators (reducing price volatility when uncertainty is high), distributed settlement parameters are generated to calculate business revenue.

[0038] Based on the aforementioned risk panorama validation set, all trained basic combinatorial models are evaluated in the three-dimensional evaluation space to identify all non-dominated solutions and form a Pareto front. In this embodiment, the three-dimensional metrics of all trained base combination models (assuming a total of M models) are organized into an evaluation matrix X, where each row represents a model and each column represents a dimension: The evaluation result corresponding to the Mth model is: ( , , ).

[0039] For each model i, iterate through all other models j ( Determine whether "model j dominates model i": Definition of dominance relationship: If and and If model j dominates model i, and model i is a “dominated solution”, then model j is excluded from the Pareto front. If no model j can dominate model i, then model i is a non-dominated solution and is retained to the Pareto front. All non-dominated solutions are sorted according to any dimension (such as policy value) to form the Pareto front.

[0040] Calculate the proportion of high-risk samples in the risk panorama verification set and use it as the corresponding risk exposure coefficient; Based on the risk exposure coefficient, the coordinates of the reference point are dynamically determined in the three-dimensional assessment space. In this embodiment, the reference point is the "ideal model" indicator of business expectations, and the ideal values ​​of each dimension need to be adjusted in conjunction with the risk exposure coefficient. The formula is as follows: In the formula, To achieve the ideal prediction accuracy, all models in the Pareto front are considered. The maximum value, This represents the ideal value for risk robustness, which is dynamically adjusted based on the risk exposure coefficient. The larger the risk exposure coefficient, the higher the requirement for robustness. The ideal value of the strategy is dynamically adjusted based on the risk exposure coefficient. The higher the risk exposure coefficient, the higher the requirement for the strategy value.

[0041] Calculate the weighted Chebyshev distance from each fundamental composite model on the Pareto front to the reference point, and select the fundamental composite model with the smallest weighted Chebyshev distance as the model with the best performance.

[0042] The formula for calculating the weighted Chebyshev distance is as follows: ; In the formula, For the i-th model in the Pareto front to the reference point The weighted Chebyshev distance, Let d be the weight of the d-th dimension. Let be the absolute deviation of model i from the ideal value in dimension d.

[0043] The predicted demand value of the transportation orders and the uncertainty quantification index are input into the strategy to generate a cellular automaton. By dynamically adjusting the state transmission resistance between cells, the supply and demand propagation is simulated, and non-uniform distributed settlement parameter adjustment instructions are generated. It should be noted that the step of inputting the predicted demand value of the transportation order and the uncertainty quantification index into the strategy to generate a cellular automaton, and simulating supply and demand propagation by dynamically adjusting the state transmission resistance between cells to generate non-uniform distributed settlement parameter adjustment instructions, includes: The target area is divided into multiple interconnected cellular grids, each representing a geographical sub-region. A state vector is maintained, including the current supply-demand ratio, historical settlement prices, driver density, and passenger waiting time. In this embodiment, driver density is innovatively subdivided into empty driver density and passenger-carrying driver density, calculated using the following formulas: Empty driver density = Number of empty drivers in a cell / Cell area; Passenger-carrying driver density = Number of passenger-carrying drivers in a cell / Cell area. The empty driver density directly determines the current capacity supply potential, providing a precise basis for adjusting settlement prices.

[0044] Based on the predicted demand for transportation orders, each cell grid interacts with its neighboring cells according to preset neighbor interaction rules, simulating the dynamic propagation process of supply and demand. In this embodiment, compared to the traditional one-way transmission of cell states, this embodiment simulates the real supply and demand flow process through two-way interaction rules of demand overflow and capacity overflow. Specifically: Demand overflow: When the predicted demand value of a cell... ( This is the average demand for the same period in history. (The corresponding multiple) and the current supply-demand ratio ( Demand overflow is triggered when the supply-demand ratio threshold is preset; otherwise, demand overflow is not triggered.

[0045] Transmission logic: Overflow The spatial decay factor decreases as the distance to adjacent cells increases, and the overflow demand is calculated based on the "capacity idle rate" of adjacent cells. The proportion of transportation capacity is allocated to ensure that demand flows to areas with sufficient capacity.

[0046] Capacity overflow: When the predicted demand value of a cell ( This is the average demand for the same period in history. (as a corresponding multiple) and empty driver density Greater than the preset multiple Historical density of empty drivers When the above conditions are met, capacity overflow is triggered; when the above conditions are not met, capacity overflow is not triggered.

[0047] Transmission logic: Overflow The driver mobility willingness coefficient is calculated based on the "price attractiveness index" (settlement price of adjacent cells / settlement price of current cells) of adjacent cells (the higher the index, the higher the coefficient, such as 1.0 when the index is 1.2 and 0.5 when the index is 0.8), ensuring that transport capacity flows to areas with high price attractiveness.

[0048] The resistance coefficient for state transfer between cell grids is dynamically adjusted using uncertainty quantification indicators. After multiple iterations, the stable states of each cell grid are collected, and a non-uniform distributed settlement parameter adjustment strategy is generated based on the principle of spatial consistency. In this way, each sub-region within different target regions obtains differentiated distributed settlement parameter adjustment instructions.

[0049] Based on the distributed settlement parameter adjustment instruction, the settlement parameter configuration results of each sub-region within the target area are directly output.

[0050] like Figure 2 As shown, this embodiment also discloses a scenario demand peak and valley prediction system based on a spatiotemporal large model. The system applies the aforementioned scenario demand peak and valley prediction method based on a spatiotemporal large model, including: The acquisition and preprocessing module is used to acquire multi-source heterogeneous spatiotemporal data of the target area, perform non-uniform spatial grid division based on the weighted Voronoi diagram algorithm, calculate the instantaneous supply-demand ratio and settlement price rolling volatility, and construct a settlement risk-enhanced spatiotemporal feature tensor. The peak-valley prediction module is used to input the settlement risk-enhanced spatiotemporal feature tensor into the pre-trained peak-valley prediction spatiotemporal model, extract features through a spatiotemporal graph attention network that integrates settlement price fluctuation features, and output the predicted value of transportation order demand and the corresponding uncertainty quantification index by a dual-branch decoder. The adjustment instruction module is used to input the predicted value of the transportation order demand and the uncertainty quantification index into the strategy to generate a cellular automaton. By dynamically adjusting the state transmission resistance between cells to simulate supply and demand propagation, it generates non-uniform distributed settlement parameter adjustment instructions. The calculation parameter output module is used to directly output the settlement parameter configuration results of each sub-region within the target area based on the distributed settlement parameter adjustment instruction.

[0051] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0052] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0053] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more electronic devices to execute all or part of the steps of the methods described in the various embodiments of this application.

[0054] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0055] In the several embodiments provided in this application, it should be understood that the disclosed application can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0056] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0057] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0058] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for predicting peak and valley demand in a scenario based on a spatiotemporal large model, characterized in that, The method includes: Acquire multi-source heterogeneous spatiotemporal data of the target region, based on weighted... The graph algorithm is used to divide the space into non-uniform grids, calculate the instantaneous supply-demand ratio and the rolling volatility of the settlement price, and construct a settlement risk-enhanced spatiotemporal feature tensor. The settlement risk-enhanced spatiotemporal feature tensor is input into a pre-trained peak-valley prediction spatiotemporal model. Features are extracted by a spatiotemporal graph attention network that integrates settlement price fluctuation features. The dual-branch decoder outputs the predicted value of transportation order demand and the corresponding uncertainty quantification index, respectively. The predicted demand value of the transportation orders and the uncertainty quantification index are input into the strategy to generate a cellular automaton. By dynamically adjusting the state transmission resistance between cells, the supply and demand propagation is simulated, and non-uniform distributed settlement parameter adjustment instructions are generated. Based on the distributed settlement parameter adjustment instruction, the settlement parameter configuration results of each sub-region within the target area are directly output.

2. The method for predicting peak and valley demand based on a spatiotemporal large model according to claim 1, characterized in that, The multi-source heterogeneous spatiotemporal data includes historical transportation order data, historical settlement price data, and real-time traffic status data.

3. The method for predicting peak and valley demand based on a spatiotemporal large model according to claim 2, characterized in that, The weighted The graph algorithm performs non-uniform spatial grid partitioning, calculates the instantaneous supply-demand ratio and settlement price rolling volatility, and constructs a settlement risk-enhanced spatiotemporal feature tensor, including: Based on the historical transportation order data and historical settlement price data, the corresponding historical order density distribution and historical settlement price volatility distribution are calculated respectively. Through weighting The graph algorithm divides the target region into a non-uniform spatial grid to generate various dynamic spatial units. Centered on the actual settlement time of historical transportation orders, a preset time window is traced back to extract the multi-source heterogeneous spatiotemporal data, forming a time-series data slice based on the settlement event; Within each of the dynamic spatial units, for each of the time-series data slices, higher-order features, including at least the instantaneous supply-demand ratio and the rolling volatility of the settlement price, are calculated; The dynamic spatial units, time-series data slices, and higher-order features are integrated to form a corresponding settlement risk-enhanced spatiotemporal feature tensor.

4. The method for predicting peak and valley demand based on a spatiotemporal large model according to claim 3, characterized in that, The training process of the pre-trained spatiotemporal model for peak-valley prediction includes: Based on a pre-set attention mechanism library and decoder structure library, the attention mechanisms in the attention mechanism library and the decoders in the decoder structure library are paired up to form basic combination models of different categories. Each basic combination model contains a spatiotemporal feature decoupling module, which is used to separate static spatial features and dynamic temporal features through a gating mechanism. Based on the risk sensitivity coefficient, N different hyperparameter configurations are dynamically generated for each basic combination model; Retrieve multi-source heterogeneous spatiotemporal historical data corresponding to the target region in the database, preprocess the multi-source heterogeneous spatiotemporal historical data to obtain the corresponding settlement risk enhanced spatiotemporal feature historical tensor, and label the settlement risk enhanced spatiotemporal feature historical tensor. Based on the volatility of historical settlement prices, the business risk value of each time segment in the sequence of the historical tensor of the settlement risk-enhanced spatiotemporal characteristics is calculated, and all historical time segments are divided into multiple risk levels according to the distribution of business risk values. From each risk level, time segments are randomly selected according to a preset ratio to form a risk panorama verification set; the remaining time segments form a risk course training set. The time segments in the risk course training set are sorted from low to high according to their business risk values ​​to construct a training course with increasing risk. According to the training course order of increasing risk, the sorted time segments are sequentially input into the basic combination model for training, and the trained basic combination models are output; wherein, the spatiotemporal feature decoupling module dynamically adjusts the fusion weight of static spatial features and dynamic temporal features in its gating mechanism according to the business risk value of the currently input time segment. The trained basic combination models are evaluated on the risk panorama validation set, and the model with the best performance is selected as the pre-trained spatiotemporal large model for peak and valley prediction.

5. The method for predicting peak and valley demand based on a spatiotemporal large model according to claim 4, characterized in that, The training courses are arranged in ascending order of risk. The ordered time segments are sequentially input into the basic combination model for training, and the trained basic combination models are output, including: Based on the training course sequence with increasing risk, the sorted time segments are divided into multiple consecutive course stages, each course stage containing a set of time segments with increasing risk values; For each time segment in each course stage, the corresponding business risk value is read, and the business risk value is concatenated with the embedding code of the current course stage. The result is input into the gated neural network, and the output is the fusion weight. Based on the fusion weight, the static spatial features and dynamic temporal features are weighted and fused to form the fusion features. The fused features are input into the encoder and the dual-branch decoder to obtain the demand forecast and the uncertainty quantification index, respectively. Calculate the loss of the demand forecasting branch and the loss of the uncertainty quantification branch, and dynamically adjust the loss weights according to the current course stage to obtain the total loss. The gradient is calculated based on the total loss, and the model parameters in the basic combined model are updated. Only when the prediction error and uncertainty calibration reach the stage advancement threshold at the same time will the model enter the next course stage. After all course stages are completed, the corresponding basic combined model will be the trained basic combined model.

6. The method for predicting peak and valley demand based on a spatiotemporal large model according to claim 5, characterized in that, The selection of the best-performing model as the pre-trained spatiotemporal large-scale peak-valley prediction model includes: A three-dimensional evaluation space is constructed, the dimensions of which include prediction accuracy dimension, risk robustness dimension, and strategy value dimension; Based on the aforementioned risk panorama validation set, all trained basic combinatorial models are evaluated in the three-dimensional evaluation space to identify all non-dominated solutions and form a Pareto front. Calculate the proportion of high-risk samples in the risk panorama verification set and use it as the corresponding risk exposure coefficient; Based on the risk exposure coefficient, the coordinates of the reference point are dynamically determined in the three-dimensional assessment space: Calculate the weighted Chebyshev distance from each fundamental composite model on the Pareto front to the reference point, and select the fundamental composite model with the smallest weighted Chebyshev distance as the model with the best performance.

7. The method for predicting peak and valley demand based on a spatiotemporal large model according to claim 6, characterized in that, The step of inputting the predicted demand value of the transportation order and the uncertainty quantification index into the strategy to generate a cellular automaton, simulating supply and demand propagation by dynamically adjusting the state transmission resistance between cells, and generating non-uniform distributed settlement parameter adjustment instructions includes: The target area is divided into multiple interconnected cellular grids, each cellular grid representing a geographic sub-region, and a state vector is maintained, which includes the current supply-demand ratio, historical settlement price, driver density, and passenger waiting time. Based on the predicted demand for transportation orders, each cell grid interacts with its neighboring cell grids according to preset neighbor interaction rules to simulate the dynamic propagation process of supply and demand. The resistance coefficient for state transfer between cell grids is dynamically adjusted using uncertainty quantification indicators. After multiple iterations, the stable states of each cell grid are collected, and a non-uniform distributed settlement parameter adjustment strategy is generated based on the principle of spatial consistency. In this way, each sub-region within different target areas obtains differentiated distributed settlement parameter adjustment instructions.

8. A scenario demand peak and valley prediction system based on a spatiotemporal large model, implementing the scenario demand peak and valley prediction method based on a spatiotemporal large model as described in any one of claims 1 to 7, characterized in that, include: The acquisition and preprocessing module is used to acquire multi-source heterogeneous spatiotemporal data of the target area, based on weighted... The graph algorithm is used to divide the space into non-uniform grids, calculate the instantaneous supply-demand ratio and the rolling volatility of the settlement price, and construct a settlement risk-enhanced spatiotemporal feature tensor. The peak-valley prediction module is used to input the settlement risk-enhanced spatiotemporal feature tensor into the pre-trained peak-valley prediction spatiotemporal model, extract features through a spatiotemporal graph attention network that integrates settlement price fluctuation features, and output the predicted value of transportation order demand and the corresponding uncertainty quantification index by a dual-branch decoder. The adjustment instruction module is used to input the predicted value of the transportation order demand and the uncertainty quantification index into the strategy to generate a cellular automaton. By dynamically adjusting the state transmission resistance between cells to simulate supply and demand propagation, it generates non-uniform distributed settlement parameter adjustment instructions. The calculation parameter output module is used to directly output the settlement parameter configuration results of each sub-region within the target area based on the distributed settlement parameter adjustment instruction.

Images