LPG sales prediction method and device based on multi-source features and adaptive network

Abstract

The invention discloses an LPG sales volume prediction method and device based on multi-source features and an adaptive network, and aims to solve the problems that the LPG terminal sales volume is affected by multi-source data, the time sequence features are complex and the general model adaptability is poor. Basic sales volume, user behaviors, external influences and operation response levels are covered; time sequence data time-frequency domain enhancement is realized through Morlet wavelet transform, and capture of an LPG service period is enhanced in combination with a service awareness position coding enhancement model; secondly, designing a multi-channel convolution-attention-cycle hybrid network MCA-RNet, and processing local fluctuations, periodic trends and external factors in parallel; and finally, outputting a prediction result through business rule verification and a dynamic optimization mechanism. According to the method, high-precision prediction of the LPG sales volume is realized, and reliable support is provided for filling operation plans, inventory optimization and dynamic pricing.

Description

Technical Field

[0001] This invention relates to the field of LPG terminal sales forecasting and intelligent decision-making technology, specifically to an LPG sales forecasting method and device based on multi-source features and adaptive networks. Background Technology

[0002] Against the backdrop of increasingly stringent regulations and large-scale integration in the LPG (oil and gas) terminal industry, accurate sales forecasting is the core support for gas stations to optimize resource allocation and reduce operating costs. However, LPG terminal sales forecasting faces multiple industry-specific challenges: First, the data sources for relevant influencing factors are complex. Sales are affected by multiple dimensions, including weather conditions, holiday cycles, user purchasing habits, cylinder turnover status, and policy adjustments, making it difficult for traditional models to effectively integrate multi-source data. Second, the time-series fluctuation characteristics are complex. LPG sales exhibit multiple overlapping cycles: daily (e.g., morning and evening filling peaks), weekly (e.g., increased purchases by catering users on weekends), and monthly (e.g., increased demand during the winter heating season). Furthermore, they are susceptible to sudden factors (e.g., cylinder inspections leading to a sharp drop in daily filling volume, extreme weather affecting transportation), making it difficult for general time-series models to accurately capture these fluctuations. Third, the industry's business logic is tightly bound. The unique differences in user types (different purchasing cycles for catering, industrial, and residential users), cylinder turnover constraints, and price elasticity characteristics of LPG terminals weaken the generalization ability of general forecasting models, resulting in forecasts that are out of sync with actual business needs. Fourth, there is insufficient adaptability to business dynamics. When gas stations add new user groups, adjust pricing strategies, or encounter policy changes, traditional models need to be retrained and cannot quickly respond to business changes.

[0003] Existing sales forecasting technologies lack customized solutions for the business characteristics of LPG terminal sales scenarios. Data processing, feature engineering, and model optimization lack deep integration with LPG industry logic, resulting in low forecast accuracy and poor business adaptability in LPG terminal scenarios, making it difficult to support actual operational decisions at gas stations. Summary of the Invention

[0004] The LPG sales forecasting method proposed in this invention, based on multi-source features and adaptive networks, can at least solve one of the technical problems in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting LPG sales based on multi-source features and adaptive networks includes the following steps: S100. Construct an LPG business characteristic system and classify and preprocess the collected LPG terminal full-service link data. S200: The pre-processed LPG terminal full-service link data is fused and time-series enhanced. S300: Construct an adaptive neural network MCA-RNet for LPG scenarios to process fused and time-enhanced LPG terminal full-service link data; S400 verifies and dynamically optimizes the LPG terminal's full-service link data after neural network processing based on business rules.

[0006] Furthermore, the method for constructing LPG service feature representation in step S100 of the present invention includes: A four-layer multimodal feature system is constructed, using LPG terminal full-service link data. Each layer of features is associated with LPG business logic. The structure of the four-layer multimodal features is as follows: Basic sales volume layer: Collect historical LPG filling volume of gas stations, and compile statistics on daily, weekly, and monthly figures, daily inventory balance, and filling ratio data of different sales regions as the core basic data for sales forecasting; User behavior layer: Based on customer management system data, user types are labeled, the average procurement cycle and single procurement volume of each type of user are calculated, user activity is statistically analyzed, and regular users are distinguished from users with sudden needs. External Influence Layer: Obtain daily average temperature, extreme weather types and duration from meteorological platforms, extract holiday cycles from calendar databases, and collect information on regional gas policy adjustments from industry regulatory platforms; Operational response layer: Records gas cylinder inspection dates, LPG price adjustment records, and promotional activity periods, and synchronously collects competitor price change data.

[0007] Furthermore, the method for fusing and timing-enhancing the preprocessed LPG terminal full-service link data in step S200 of the present invention includes: To address the multi-period characteristics of LPG sales time-series data, three sets of scale parameters—short, medium, and long—are preset to correspond to the feature extraction requirements of daily high-frequency fluctuations, weekly medium-term changes, and monthly low-frequency trends, respectively. The one-dimensional sales sequence is decomposed into multi-frequency domain features using Morlet wavelet transform, generating a multi-channel feature matrix containing high-frequency detail components and low-frequency approximation components. The specific processing flow is as follows: Morlet wavelet basis functions:

[0008] In the formula, For the center frequency, It is a time variable; Multiscale wavelet decomposition:

[0009] In the formula, To represent different scales, τ is the translation coefficient, and τ is the sampling interval; Wavelet transform of time-series signals:

[0010] Key event feature weighting:

[0011] Multi-frequency domain feature matrix generation:

[0012] Business-aware location coding: The time stamps unique to LPG business are converted into numerical periodic parameters and fused with the traditional Transformer location coding in a weighted manner; among them, the weight of the daily gas cylinder detection identifier is set to 1.5, the weight of the holiday identifier is set to 1.2, and the weight of the weekday identifier is set to 1.0, so that the model prioritizes the key time nodes of the LPG terminal in time series learning. Basic location encoding:

[0013]

[0014] In the formula, For time steps; For feature dimension, For index dimensions, ; Business tag value mapping:

[0015] In the formula, For indicator functions, time steps , belongs to the The value is 1 if it is a business category tag, otherwise it is 0; Weighted fusion location coding:

[0016] Feature fusion: .

[0017] Furthermore, the method for constructing the LPG scene adaptive neural network MCA-RNet in step S300 of the present invention includes: Design a multi-path convolutional-attention-recurrent hybrid network to process different feature dimensions of LPG sales in parallel, wherein: The multi-path structure design is as follows: Local fluctuation feature extraction path: adopts a multi-scale CNN architecture, configured with two small-sized convolutional kernels of 2×2 and 3×3, to capture the short-term fluctuation details of daily sales, and retains key local features through ReLU activation function and max pooling layer; Convolution operation:

[0018]

[0019] For the input feature map, , The kernels are 2×2 and 3×3, respectively. ReLU activation and max pooling:

[0020]

[0021] In the formula, Max pooling with a window size of 2 and a stride of 2; Periodic trend feature extraction pathway: Using an LSTM network with 3 hidden layers, the system uses the synergistic effect of the forget gate, input gate, and output gate to memorize the weekly and monthly sales change patterns, thus solving the problem of long sequence dependencies.

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] In the formula, For the first Step input, , These are the hidden state and the cellular state, respectively. , This is the weight matrix.

[0028] Furthermore, the multi-path convolutional-attention-recurrent hybrid network of the present invention also includes: External correlation feature extraction path: The Transformer encoder is used to calculate the correlation weight between external factors and sales through a self-attention mechanism, so as to achieve accurate correlation of multi-dimensional features; Multi-head attention and feedforward networks:

[0029]

[0030]

[0031]

[0032] For input embedding, W Q W V W K For the projection matrix, For attention head dimension; 6-layer encoder output:

[0033] Feature fusion and prediction: The output features from the three paths enter the attention fusion module, and the weights are dynamically allocated according to the LPG business scenario; Dynamic weight allocation:

[0034] Fully connected layers and inverse transform:

[0035]

[0036] The model is iteratively trained on the training set using an appropriate optimizer and loss function. After each training round, the performance is evaluated using a validation set. The model convergence effect is optimized by adjusting hyperparameters such as the learning rate and dropout probability until the error index of the model on the validation set converges to a preset threshold, thus completing the training of the MCA-RNet neural network.

[0037] Furthermore, the method for verifying and dynamically optimizing the LPG terminal full-service link data in step S400 of the present invention includes: A rule base for LPG terminal services has been constructed, comprising three core rule categories: First, the predicted sales volume must not exceed the current total inventory of gas cylinders to avoid over-inventory forecasts. Second, the daily sales fluctuation range does not exceed ±30% of the historical average for the same period, filtering out abnormal predictions. Third, sales changes before and after holidays must conform to users' purchasing habits; For predictions that violate the rules, corrections are made based on the sales patterns of similar historical scenarios.

[0038] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0039] In summary, this solution is deeply integrated with LPG terminal business in all technical aspects, from feature system construction to business rule verification, solving the problem of the disconnect between general models and industry logic. The prediction results can directly support actual operational decisions and enhance scenario adaptability. By using Morlet wavelet transform time-series enhancement and MCA-RNet multi-path network, the multi-period fluctuations and multi-factor correlations of LPG sales can be effectively captured. In actual LPG terminal data testing, compared with the general CNN-Transformer model, RMSE is reduced by 15%-20%, MAE is reduced by 12%-18%, and the prediction error in special scenarios can be controlled within 5.5%, thus improving prediction accuracy. The incremental learning mechanism enables lightweight model updates, completing parameter adjustments within 24 hours of adding new data without retraining, and can quickly adapt to business changes such as gas station user growth, pricing strategy adjustments, and policy changes. Attached Figure Description

[0040] Figure 1 This is an overall flowchart of the LPG terminal sales forecasting method of the present invention; Figure 2 This is a schematic diagram of the MCA-RNet multipath network structure of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0042] like Figure 1 As shown in this embodiment, the LPG sales prediction method based on multi-source features and adaptive networks includes the following steps: S100. Construct an LPG business characteristic system and classify and preprocess the collected LPG terminal full-service link data. S200: The pre-processed LPG terminal full-service link data is fused and time-series enhanced. S300: Construct an adaptive neural network MCA-RNet for LPG scenarios to process fused and time-enhanced LPG terminal full-service link data; S400: Verify and dynamically optimize the LPG terminal's full-service link data after neural network processing based on business rules; The following provides a detailed explanation of each step: S100. Construct an LPG business characteristic system and classify and preprocess the collected LPG terminal full-service link data. A four-layer multimodal feature system is constructed around the entire LPG terminal sales business process to collect data from the entire LPG terminal business chain. Each layer of features is strongly correlated with the LPG business logic. The specific structure of the four-layer multimodal feature system is as follows: Basic sales volume layer: Collect historical LPG filling volume of gas stations (statistics by day, week, and month), daily inventory balance, and filling ratio data of different sales regions as the core basic data for sales forecasting; User behavior layer: Based on customer management system data, user types are labeled (catering, business, residents), the average purchase cycle and single purchase volume of each type of user are calculated, user activity is statistically analyzed (such as the number of purchases in the past 30 days), and regular users are distinguished from users with sudden needs (such as users handling business). External Influence Layer: Obtain daily average temperature, extreme weather types and duration from meteorological platforms, extract holiday cycles from calendar databases, and collect information on regional gas policy adjustments from industry regulatory platforms; Operational response layer: Records cylinder inspection dates, LPG price adjustment records, and promotional activity periods, and simultaneously collects competitor price change data.

[0043] The data collected from the above four layers are classified by layer to form a structured raw dataset, and then integrated in a structured manner. A multimodal dataset is formed by aligning the data with timestamps, and each data point is labeled with the corresponding LPG business scenario label (such as "winter heating season", "before the Spring Festival", "gas cylinder inspection day").

[0044] Data standardization is performed using "3σ+business rules," employing a dual mechanism to remove outlier data. Abnormal sales data for special scenarios such as cylinder inspection days and equipment failure days are marked and filtered. Missing data is filled in according to scenario-adaptive strategies: for missing time dimensions, the nearest similar date interpolation method is used, and for missing spatial dimensions, the filler value is generated based on the correlation features of similar users or regions. Numerical features are normalized using Min-Max, and categorical features are one-hot encoded, uniformly mapping the data to the format of the model input. Training, validation, and test sets are divided according to a preset ratio, and each dataset is labeled with LPG business scenario tags to complete data preprocessing.

[0045] S200: The pre-processed LPG terminal full-service link data is fused and time-series enhanced. Morlet wavelet transform time-series enhancement: Addressing the multi-period characteristics of LPG sales time-series data, three sets of scale parameters (short, medium, and long) are preset to correspond to the feature extraction needs of daily high-frequency fluctuations, weekly mid-term changes, and monthly low-frequency trends, respectively. The one-dimensional sales sequence is decomposed into multi-frequency domain features through Morlet wavelet transform, generating a multi-channel feature matrix containing high-frequency detail components and low-frequency approximation components. The weight of feature components corresponding to extreme weather and holidays is increased by 30%-50%, strengthening the impact of key events; abnormally low sales noise generated on cylinder inspection days is filtered out to ensure feature quality.

[0046] 1. Morlet wavelet transform timing enhancement (1) Morlet wavelet basis functions

[0047] In the formula, For the center frequency, It is a time variable.

[0048] (2) Multiscale wavelet decomposition

[0049] In the formula, To represent different scales, τ is the translation coefficient, and τ is the sampling interval.

[0050] (3) Wavelet transform of time-series signals

[0051] (4) Weighted analysis of key event characteristics

[0052] (5) Generation of multi-frequency domain feature matrix

[0053] Business-aware location coding: The time stamps specific to LPG business are converted into numerical periodic parameters and then weighted and fused with the traditional Transformer location coding. Specifically, the weight of the daily cylinder inspection identifier is set to 1.5, the weight of the holiday identifier is set to 1.2, and the weight of the weekday identifier is set to 1.0. This allows the model to prioritize key time nodes of the LPG terminal during time-series learning, improving its ability to perceive the business cycle.

[0054] (1) Basic location coding

[0055]

[0056] In the formula, For time steps; For feature dimension, For index dimensions, .

[0057] (2) Business tag numerical mapping

[0058] In the formula, For indicator functions, time steps , belongs to the The value is 1 if it is a business category label, and 0 otherwise.

[0059] (3) Weighted fusion location coding

[0060] (4) Feature fusion

[0061] S300: Construct an adaptive neural network MCA-RNet for LPG scenarios to process fused and time-enhanced LPG terminal full-service link data; Design a multi-path convolutional-attention-recurrent hybrid network to process different feature dimensions of LPG sales in parallel, including: The multi-path structure design is as follows: Local fluctuation feature extraction path: adopts a multi-scale CNN architecture, configured with two small-sized convolutional kernels of 2×2 and 3×3, to capture the short-term fluctuation details of daily sales, and retains key local features through ReLU activation function and max pooling layer; (1) Convolution operation

[0062]

[0063] For the input feature map, , The kernels are 2×2 and 3×3, respectively.

[0064] (2) ReLU activation and max pooling

[0065]

[0066] In the formula, Max pooling with a window size of 2 and a step size of 2.

[0067] Periodic trend feature extraction pathway: Using an LSTM network with 3 hidden layers, the system uses the synergistic effect of the forget gate, input gate, and output gate to memorize the weekly and monthly sales change patterns, thus solving the problem of long sequence dependencies.

[0068]

[0069]

[0070]

[0071]

[0072]

[0073] In the formula, For the first Step input, , These are the hidden state and the cellular state, respectively. , This is the weight matrix. For bias vectors, For the Sigmod function, For the hidden layer dimension.

[0074] External correlation feature extraction path: The Transformer encoder (6-layer encoder) is used to calculate the correlation weight between external factors and sales through a self-attention mechanism, highlighting key influencing factors, weakening the interference of irrelevant factors, and achieving accurate correlation of multi-dimensional features; (1) Multi-head attention and feedforward networks

[0075]

[0076]

[0077]

[0078] For input embedding, W Q W V W K For the projection matrix, For attention head dimension.

[0079] (2) Output of 6-layer encoder

[0080] Feature fusion and prediction: The output features from the three pathways enter the attention fusion module, and the weights are dynamically allocated according to the LPG business scenario. For example, the weight of cyclical trend features is increased during winter, and the weight of external correlation features is increased during holidays. The fused features are input into the fully connected layer, and the normalized prediction value is output through the activation function. Then, the actual sales prediction result is obtained through the business scale inverse transformation.

[0081] (1) Dynamic weight allocation

[0082] (2) Fully connected layers and inverse transform

[0083]

[0084] The model is iteratively trained on the training set using an appropriate optimizer and loss function. After each training round, the performance is evaluated using a validation set. The model convergence effect is optimized by adjusting hyperparameters such as the learning rate and dropout probability until the error index of the model on the validation set converges to a preset threshold, thus completing the training of the MCA-RNet neural network.

[0085] S400: Verify and dynamically optimize the LPG terminal's full-service link data after neural network processing based on business rules; A rule base for LPG terminal business constraints was constructed, comprising three core rules: First, predicted sales must not exceed the current total cylinder inventory to avoid "over-inventory predictions"; second, daily sales fluctuations must not exceed ±30% of the historical average for the same period to filter out abnormal predictions; and third, sales changes before and after holidays must conform to user purchasing habits (e.g., sales increase 7 days before the Spring Festival and decrease 5 days after). Prediction results that violate these rules are corrected based on historical patterns of similar scenarios (e.g., sales changes around the Spring Festival in the past 3 years) to ensure that predicted values ​​align with actual operational logic. (1) Inventory constraints

[0086] These are the initial predicted values. This represents the current inventory.

[0087] (2) Fluctuation constraint

[0088] Historical average sales for the same period Correct sales figures.

[0089] (3) Holiday restrictions

[0090] Adjustment factor for holidays: before holidays After the holiday .

[0091] Incremental learning and dynamic optimization: A sliding time window is used to select the latest LPG business data. Only the fully connected layer parameters and attention weights of MCA-RNet are updated, without reconstructing the overall network structure, thus reducing the model update cost. A lightweight update is triggered every week of new data. After the update, the model performance is verified using real sales data from the past two weeks. If the prediction error (MAE) exceeds the threshold, the hidden layer parameters of the network are fine-tuned again to ensure that the model continuously adapts to business changes. This enables the model to dynamically adapt to business changes (such as new user groups or policy adjustments) and avoids a decrease in prediction accuracy due to data lag.

[0092] Model performance monitoring and iterative optimization: Real-time performance monitoring establishes a model performance monitoring system to track key parameters such as error metrics between the training and validation sets, feature weight distribution of each pathway, and deviation rate between prediction results and actual sales in real time. Performance thresholds are set for different business scenarios of LPG. When the monitoring metrics in a certain scenario exceed the threshold, an early warning mechanism is triggered, prompting that model optimization is required.

[0093] Iterative optimization and model supplementation: When monitoring detects a decline in model performance, re-optimize the model hyperparameters (convolution kernel size, number of LSTM hidden layer nodes, number of attention heads), or adjust the feature fusion weights and business rule verification thresholds; if the existing model structure cannot adapt to new business scenarios, supplement the MCA-RNet with sub-modules adapted to the new scenarios, or expand the basic model pool to ensure that the prediction method continues to meet the needs of LPG terminal business development and scenario changes, and maintain the reliability and advancement of the model.

[0094] In summary, this solution is deeply integrated with LPG terminal business in all technical aspects, from feature system construction to business rule verification, solving the problem of the disconnect between general models and industry logic. The prediction results can directly support actual operational decisions and enhance scenario adaptability. By using Morlet wavelet transform time-series enhancement and MCA-RNet multi-path network, the multi-period fluctuations and multi-factor correlations of LPG sales can be effectively captured. In actual LPG terminal data testing, compared with the general CNN-Transformer model, RMSE is reduced by 15%-20%, MAE is reduced by 12%-18%, and the prediction error in special scenarios can be controlled within 5.5%, thus improving prediction accuracy. The incremental learning mechanism enables lightweight model updates, completing parameter adjustments within 24 hours of adding new data without retraining, and can quickly adapt to business changes such as gas station user growth, pricing strategy adjustments, and policy changes.

[0095] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the LPG sales forecasting methods based on multi-source features and adaptive networks described in the above embodiments.

[0096] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0097] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0098] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0099] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0100] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting LPG sales based on multi-source features and adaptive networks, characterized in that, Includes the following steps: S100. Construct an LPG business characteristic system and classify and preprocess the collected LPG terminal full-service link data. S200: The pre-processed LPG terminal full-service link data is fused and time-series enhanced. S300: Construct an adaptive neural network MCA-RNet for LPG scenarios to process fused and time-enhanced LPG terminal full-service link data; S400 verifies and dynamically optimizes the LPG terminal's full-service link data after neural network processing based on business rules.

2. The LPG sales forecasting method based on multi-source features and adaptive networks according to claim 1, characterized in that, The method for constructing the LPG business characteristic representation in step S100 includes: A four-layer multimodal feature system is constructed, using LPG terminal full-service link data. Each layer of features is associated with LPG business logic. The structure of the four-layer multimodal features is as follows: Basic sales volume layer: Collect historical LPG filling volume of gas stations, and compile statistics on daily, weekly, and monthly figures, daily inventory balance, and filling ratio data of different sales regions as the core basic data for sales forecasting; User behavior layer: Based on customer management system data, user types are labeled, the average procurement cycle and single procurement volume of each type of user are calculated, user activity is statistically analyzed, and regular users are distinguished from users with sudden needs. External Influence Layer: Obtain daily average temperature, extreme weather types and duration from meteorological platforms, extract holiday cycles from calendar databases, and collect information on regional gas policy adjustments from industry regulatory platforms; Operational response layer: Records gas cylinder inspection dates, LPG price adjustment records, and promotional activity periods, and synchronously collects competitor price change data.

3. The LPG sales forecasting method based on multi-source features and adaptive networks according to claim 1, characterized in that, Step S200 involves fusing and timing-enhancing the preprocessed LPG terminal full-service link data, including the following methods: To address the multi-period characteristics of LPG sales time-series data, three sets of scale parameters—short, medium, and long—are preset to correspond to the feature extraction requirements of daily high-frequency fluctuations, weekly mid-term changes, and monthly low-frequency trends, respectively. The one-dimensional sales sequence is decomposed into multi-frequency domain features using Morlet wavelet transform, generating a multi-channel feature matrix containing high-frequency detail components and low-frequency approximation components. The processing flow is as follows: Morlet wavelet basis functions: In the formula, For the center frequency, It is a time variable; Multiscale wavelet decomposition: In the formula, To represent different scales, τ is the translation coefficient, and τ is the sampling interval; Wavelet transform of time-series signals: Key event feature weighting: Multi-frequency domain feature matrix generation: Business-aware location coding: Convert the time stamps unique to LPG business into numerical periodic parameters and perform weighted fusion with Transformer's traditional location coding, so that the model prioritizes key time nodes of LPG terminals in time-series learning; Basic location encoding: In the formula, For time steps; For feature dimension, For index dimensions, ; Business tag value mapping: In the formula, For indicator functions, time steps , belongs to the The value is 1 if it is a business category tag, otherwise it is 0; Weighted fusion location coding: Feature fusion: 。 4. The LPG sales forecasting method based on multi-source features and adaptive networks according to claim 1, characterized in that, The method for constructing the LPG scene adaptive neural network MCA-RNet in step S300 includes: Design a multi-path convolutional-attention-recurrent hybrid network to process different feature dimensions of LPG sales in parallel, wherein: The multi-path structure design is as follows: Local fluctuation feature extraction path: adopts a multi-scale CNN architecture, configured with two small-sized convolutional kernels of 2×2 and 3×3, to capture the short-term fluctuation details of daily sales, and retains key local features through ReLU activation function and max pooling layer; Convolution operation: For the input feature map, , The kernels are 2×2 and 3×3, respectively. ReLU activation and max pooling: In the formula, Max pooling with a window size of 2 and a stride of 2; Periodic trend feature extraction pathway: Using an LSTM network with 3 hidden layers, the system uses the synergistic effect of the forget gate, input gate, and output gate to memorize the weekly and monthly sales change patterns, thus solving the problem of long sequence dependencies. In the formula, For the first Step input, , These are the hidden state and the cellular state, respectively. , This is the weight matrix.

5. The LPG sales forecasting method based on multi-source features and adaptive networks according to claim 4, characterized in that, The multi-path convolutional-attention-recurrent hybrid network also includes: External correlation feature extraction path: The Transformer encoder is used to calculate the correlation weight between external factors and sales through a self-attention mechanism, so as to achieve accurate correlation of multi-dimensional features; Multi-head attention and feedforward networks: For input embedding, W Q W V W K For the projection matrix, For attention head dimension; 6-layer encoder output: Feature fusion and prediction: The output features from the three paths enter the attention fusion module, and the weights are dynamically allocated according to the LPG business scenario; Dynamic weight allocation: Fully connected layers and inverse transform: The model is iteratively trained on the training set using an appropriate optimizer and loss function. After each training round, the performance is evaluated using a validation set. The model convergence effect is optimized by adjusting hyperparameters such as the learning rate and dropout probability until the error index of the model on the validation set converges to a preset threshold, thus completing the training of the MCA-RNet neural network.

6. The LPG sales forecasting method based on multi-source features and adaptive networks according to claim 1, characterized in that, The method for verifying and dynamically optimizing the LPG terminal's full-service link data in step S400 includes: A rule base for LPG terminal services has been constructed, comprising three core rule categories: First, the predicted sales volume must not exceed the current total inventory of gas cylinders to avoid over-inventory forecasts. Second, the daily sales fluctuation range does not exceed ±30% of the historical average for the same period, filtering out abnormal predictions. Third, sales changes before and after holidays must conform to users' purchasing habits; For predictions that violate the rules, corrections are made based on the sales patterns of similar historical scenarios.

7. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program that, when executed by the processor, causes the processor to perform the method as described in any one of claims 1 to 6.

Images