Improved LSTM (Long Short Term Memory)-based space division fast sale product delivery strategy method
Through the improved LSTM network combining CNN and ResNet to build a CRL-Net model, the complexity of fast-selling product price prediction is solved, accurate prediction of future price trends is achieved, and the company's market response capabilities and operating efficiency are improved.
Patent Information
- Application Number
- CN202510326566.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-19
- Publication Date
- 2025-07-01
AI Technical Summary
The existing fast-selling product price prediction model is difficult to accurately capture complex market price trends, especially in the face of nonlinear, multivariable and long-term dependency characteristics, and lacks generalization capabilities, resulting in large prediction deviations and poor adaptability, which affects inventory management and pricing strategies.
The improved LSTM network is used to combine convolutional neural network (CNN) and residual connection (ResNet), and through feature extraction and data augmentation, a CRL-Net model is constructed to capture the long-term dependence of fast-selling product prices, and a spatial zoning delivery strategy is generated through anti-normalization processing.
It significantly improves the accuracy and practicality of fast-selling product price prediction, helps enterprises optimize inventory management and pricing strategies, and improve operating efficiency.
Smart Images
Figure CN120235637A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of industrial big data processing, and particularly relates to a method for fast-moving consumer goods placement strategy in spatial zoning based on improved LSTM. Background Art
[0002] The core technical challenge in fast-moving consumer goods price prediction lies in how to accurately capture and predict highly complex and frequently fluctuating market price trends. The fast-moving consumer goods market is affected by various factors, such as seasonal demand, raw material costs, competitor strategies, etc., resulting in price characteristics such as non-linearity, multi-variables, long-term and short-term dependencies. Traditional statistical models are difficult to effectively handle such complex price change patterns, often showing problems such as large prediction deviations and poor adaptability. This directly affects the enterprise's inventory management, pricing strategy and profit level. At the same time, there are a large number of fast-moving consumer goods categories, and the price correlation between different categories also poses challenges to modeling. How to extract key features from massive historical data, construct a model architecture that can capture both short-term price fluctuations and long-term trends, and on this basis achieve accurate prediction of future price trends is a technical problem to be solved urgently. In addition, the model needs to have strong generalization ability to adapt to price prediction tasks in different categories and different market environments. Solving this technical problem requires not only innovation at the algorithm level, but also in-depth understanding of the characteristics of the fast-moving consumer goods industry and the organic combination of domain knowledge and advanced artificial intelligence technology. Summary of the Invention
[0003] The present invention provides a method for fast-moving consumer goods placement strategy in spatial zoning based on improved LSTM, which mainly includes: obtaining historical price data of fast-moving consumer goods and dividing it into a test set and a training set; inputting the data in the training set into the CRL-Net fast-moving consumer goods price prediction model for data training, and the CRL-Net fast-moving consumer goods price prediction model is constructed by combining a convolutional neural network CNN, a residual connection ResNet and a long short-term memory LSTM network; inputting the data in the test set into the trained CRL-Net fast-moving consumer goods price prediction model, using the convolutional neural network CNN to extract features from the historical price data of the fast-moving consumer goods to obtain fast-moving consumer goods price feature data; using the residual connection ResNet to perform data enhancement on the fast-moving consumer goods price feature data; inputting the enhanced fast-moving consumer goods price feature data into the long short-term memory LSTM network to dynamically predict the price of fast-moving consumer goods to obtain a fast-moving consumer goods price prediction result, and the long short-term memory LSTM network is used to capture long-term dependencies in time series data; performing inverse normalization processing on the fast-moving consumer goods price prediction result to obtain the final fast-moving consumer goods placement strategy in spatial zoning.
[0004] Further, the feature extraction of the fast-moving consumer goods historical price data by using a convolutional neural network (CNN) includes: windowing the fast-moving consumer goods historical price data to form fast-moving consumer goods price window data; performing a convolution operation on the fast-moving consumer goods price window data to extract local feature information; and determining the fast-moving consumer goods price feature data according to the extracted local feature information.
[0005] Further, the data augmentation of the fast-moving consumer goods price feature data by using a residual connection (ResNet) includes: constructing a plurality of residual blocks, each of the residual blocks including two convolutional layers; performing a feature transformation on the fast-moving consumer goods price feature data through the residual blocks while retaining the original input features; adding the transformed features to the original input features to obtain a new feature representation form, enhancing the learning ability and stability of the model.
[0006] Further, inputting the enhanced fast-moving consumer goods price feature data into a long short-term memory (LSTM) network for training includes: controlling the degree of forgetting of the long-term state through a forgetting gate; controlling the storage degree of new input data through an input gate; controlling the degree of output of the current state to the next moment state through an output gate; and updating the long-term and short-term states of the current moment according to the input of the current moment and the state of the previous moment.
[0007] Further, the inverse normalization processing of the fast-moving consumer goods price prediction result includes: obtaining the maximum value and the minimum value of the normalization processing; calculating the actual fast-moving consumer goods price prediction value according to the predicted normalized fast-moving consumer goods historical price data in combination with the maximum value and the minimum value; and outputting the inverse-normalized fast-moving consumer goods price prediction result as the basis for the final spatial zoning placement strategy.
[0008] The technical solution provided by the embodiment of the present invention may include the following beneficial effects:
[0009] The present invention discloses a spatial zoning fast-moving consumer goods placement strategy method based on an improved LSTM. Aiming at the problems of frequent price fluctuations and complex influencing factors in the fast-moving consumer goods market, by acquiring historical price data and performing preprocessing, and combining a convolutional neural network, a residual network and a long short-term memory network, a deep learning model is constructed. The model can effectively extract the feature representation of price data and capture the long-term dependence relationship of price fluctuations, so as to achieve accurate prediction of future price trends. The present invention iteratively optimizes the model parameters, continuously improves the prediction accuracy, and finally applies the trained model to the actual business scenario, providing a reliable reference basis for fast-moving consumer goods sales and pricing decisions. This method significantly improves the accuracy and practicality of fast-moving consumer goods price prediction, helps enterprises better respond to market changes, optimize inventory management and pricing strategies, and thus improves the operating efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0010] Figure 1 Flow chart of a method for fast-moving consumer goods placement strategy based on improved LSTM for the present invention;
[0011] Figure 2 Prediction effect based on LSTM model in 2019;
[0012] Figure 3 Prediction effect based on LSTM-CNN model in 2019;
[0013] Figure 4 Prediction effect based on CRL-Net model in 2019;
[0014] Figure 5 Prediction effect based on LSTM model in 2020;
[0015] Figure 6 Prediction effect based on LSTM-CNN model in 2020;
[0016] Figure 7 Prediction effect based on CRL-Net model in 2020;
[0017] Figure 8 Prediction effect based on LSTM model in 2021;
[0018] Figure 9 Prediction effect based on LSTM-CNN model in 2021;
[0019] Figure 10 Prediction effect based on CRL-Net model in 2021;
[0020] Figure 11 Model training data;
[0021] Figure 12 Prediction data. Detailed implementation manner
[0022] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this specification without creative efforts shall fall within the scope of protection of this specification.
[0023] Such as Figure 1 , a method for fast-moving consumer goods placement strategy based on improved LSTM in this embodiment may specifically include:
[0024] S101. Obtain historical price data of fast-moving consumer goods and divide it into a test set and a training set.
[0025] The historical price data of fast-moving consumer goods includes the sales time and sales price. The data is classified and sorted according to the sales area to obtain the historical price data of fast-moving consumer goods divided by region. The historical price data of fast-moving consumer goods is cleaned and preprocessed to remove invalid data such as missing values and outliers. Through data standardization, data of different scales are converted to the same scale to facilitate subsequent analysis and modeling. The historical price data of fast-moving consumer goods is divided into training sets and test sets according to the time series. The data set division is optimized through the cross-validation method to improve the generalization ability of the model and prepare for model training and evaluation.
[0026] S102: Input the data in the training set into the CRL-Net fast-moving consumer goods price prediction model for data training.
[0027] Specifically, during the data training process, the learning rate was set to 0.01, the number of iterations was set to 500, and the mean square error loss function was minimized by the Adam optimization algorithm to complete the training of the fast-moving consumer goods price prediction model. After the training is completed, the mean absolute percentage error (MAPE) of the model is calculated using the test set data. If the MAPE is less than 5%, the model prediction performance is considered to be good and can be used to guide the subsequent spatial zoning fast-moving consumer goods placement strategy.
[0028] S103: Input the data in the test set into the trained CRL-Net fast-moving consumer goods price prediction model, and output the fast-moving consumer goods price prediction results.
[0029] Feature extraction is performed on the fast-moving consumer goods price training data using a convolutional neural network (CNN). Through operations such as convolution and pooling, local features in the data are automatically learned to obtain deep features representing the changing patterns of fast-moving consumer goods prices. Based on the features extracted by the convolutional neural network, a residual network (ResNet) structure is introduced. Through identity mapping and cross-layer connections, feature transmission is enhanced, improving the network's learning ability and stability, and capturing complex features in the historical price data of fast-moving consumer goods. The local features extracted by the CNN and the complex features learned by the ResNet are input into a long short-term memory network (LSTM). Using the memory mechanism and gating units of the LSTM, the long-term dependence relationship of the fast-moving consumer goods price time series is modeled. A loss function and an optimization algorithm are set in the LSTM network. Through the forward propagation and backpropagation processes, the model parameters are iteratively optimized to minimize the error between the predicted price and the true price, obtaining a trained fast-moving consumer goods price prediction model. The trained improved LSTM network is used to predict the fast-moving consumer goods price test set. The prediction accuracy of the model is measured through evaluation indicators such as MAE and RMSE, and the differences in prediction effects in different regions are analyzed to provide a quantitative basis for the fast-moving consumer goods placement strategy in spatial zoning.
[0030] Further, feature extraction is performed on the historical price data of the fast-moving consumer goods using a convolutional neural network CNN, including: windowing the historical price data of the fast-moving consumer goods to form fast-moving consumer goods price window data; for the fast-moving consumer goods price window data, performing a convolution operation to extract local feature information; and determining fast-moving consumer goods price feature data according to the extracted local feature information.
[0031] Further, data enhancement is performed on the fast-moving consumer goods price feature data using a residual connection ResNet, including: constructing multiple residual blocks, each residual block including two convolutional layers; for the fast-moving consumer goods price feature data, performing feature transformation through the residual blocks while retaining the original input features; and adding the transformed features to the original input features to obtain a new feature representation form.
[0032] The enhanced fast-moving consumer goods price feature data is input into a long short-term memory LSTM network to perform dynamic prediction on the fast-moving consumer goods price, including: controlling the degree of forgetting of the long-term state through a forgetting gate; controlling the storage degree of new input data through an input gate; controlling the degree of output of the current state to the next moment state through an output gate; and updating the long-term and short-term states at the current moment according to the input at the current moment and the state at the previous moment.
[0033] Specifically, the state matrix h at the previous moment (t-1) and the input vector x at time t (t) Through the action of an activation function sigmoid, the output f of the forgetting gate is obtained (t) . The output f here (t)Represents the probability of forgetting the previous hidden cell state. The mathematical expression is as follows:
[0034] f (t) = σ(W f h (t-1) + U f x (t) + b f ), where W f , U f , b f are the coefficients and biases of the linear relationship. σ is the ReLU activation function.
[0035] The input gate updates the cell state according to the input fast-moving consumer goods price data and historical information. The input gate is responsible for processing the input at the current sequence position. It consists of two parts. The first part uses the ReLU activation function, and the output is i (t) , and the second part uses the tanh activation function, and the output is a (t) . After multiplying the results of the two, the cell state is updated. The mathematical expression is as follows:
[0036] i (t) = σ(W i h (t-1) + U i x (t) + b i ), a (t) = tanh(W a h (t-1) + U a x (t) + b a ),
[0037] where W i , U i , b i , W a , U a , b a are the coefficients and biases of the linear relationship. σ is the ReLU activation function.
[0038] The output gate outputs the current fast-moving consumer goods price information according to the cell state. The update of the state matrix h (t) at the current moment consists of two parts. The first part is o (t) , which is obtained from the state matrix h (t-1) at the previous moment and the input vector x (t) at time t, and the activation function ReLU. The second part consists of the hidden state C (t) and the tanh activation function. The mathematical expression is as follows:
[0039] o (t) = σ(Wo h (t-1) +U o x (t) +b o ),h (t) =o (t) ⊙tanh(C (t) )。
[0040] S104. Denormalize the fast-moving consumer goods price prediction result to obtain the final fast-moving consumer goods placement strategy for spatial zoning.
[0041] The denormalization process is specifically as follows: Obtain the maximum and minimum values of the normalization process; According to the predicted normalized historical fast-moving consumer goods price data, combine the maximum and minimum values to calculate the actual fast-moving consumer goods price prediction value; Output the denormalized fast-moving consumer goods price prediction result.
[0042] In the present invention, the LSTM network is improved by combining CNN and ResNet. The trained CNN model is used for feature extraction to obtain the high-dimensional feature representation of the fast-moving consumer goods price. Through the residual connection structure, the extracted feature data is passed to a deeper neural network to enhance the model's ability to capture and learn complex feature patterns. The feature data extracted by CNN and the output of the residual connection are used as the input of the LSTM network, and the time series modeling ability of LSTM is used to predict the future price trend. Set a sliding time window, and obtain the latest historical fast-moving consumer goods price data at each time step as the input of the model to update the price prediction result in real time. Business rules and expert knowledge are used to post-process and correct the prediction results of the LSTM network to improve the interpretability and credibility of the prediction. The predicted price trend information is integrated into the existing sales and pricing decision-making system to assist business personnel in formulating reasonable sales plans and pricing strategies. Continuously monitor the performance of the fast-moving consumer goods price prediction model, and retrain and optimize the model regularly by collecting actual sales data and price feedback. Build a visual price prediction and decision support system to enable business personnel to conveniently query price prediction information and make dynamic adjustments to sales and pricing according to the prediction results.
[0043] Taking cigarettes, a fast-moving consumer good, as an example, the technical solution of the present invention will be illustrated. The data used is the cigarette sales data from 2011 to 2022. Among them, the cigarette product specification sales data in 2011 was 46,695 records, the cigarette product specification sales data in 2012 was 49,046 records, the cigarette product specification sales data in 2013 was 52,709 records, the cigarette product specification sales data in 2014 was 53,677 records, the cigarette product specification sales data in 2015 was 56,216 records, the cigarette product specification sales data in 2016 was 63,236 records, the cigarette product specification sales data in 2017 was 67,233 records, the cigarette product specification sales data in 2018 was 68,186 records, the cigarette product specification sales data in 2019 was 71,486 records, the cigarette product specification sales data in 2020 was 65,082 records, the cigarette product specification sales data in 2021 was 66,642 records, and each data table contains 12 columns; the cigarette sales amount per case data in 2011 was 4,436 records, the cigarette sales amount per case data in 2012 was 5,314 records, the cigarette sales amount per case data in 2013 was 5,790 records, the cigarette sales amount per case data in 2014 was 6,094 records, the cigarette sales amount per case data in 2015 was 6,412 records, the cigarette sales amount per case data in 2016 was 6,886 records, the cigarette sales amount per case data in 2017 was 7,246 records, the cigarette sales amount per case data in 2018 was 7,530 records, the cigarette sales amount per case data in 2019 was 7,754 records, the cigarette sales amount per case data in 2020 was 8,250 records, the cigarette sales amount per case data in 2021 was 8,398 records, and the cigarette sales amount per case data in 2022 was 8,504 records, and each data table contains 28 columns; the total number of cigarette sales cases in provinces in 2011 was 352 cases, the total number of cigarette sales cases in provinces in 2012 was 355 cases, the total number of cigarette sales cases in provinces in 2013 was 355 cases, the total number of cigarette sales cases in provinces in 2014 was 356 cases, the total number of cigarette sales cases in provinces in 2015 was 356 cases, the total number of cigarette sales cases in provinces in 2016 was 356 cases, the total number of cigarette sales cases in provinces in 2017 was 356 cases, the total number of cigarette sales cases in provinces in 2018 was 356 cases, the total number of cigarette sales cases in provinces in 2019 was 358 cases, the total number of cigarette sales cases in provinces in 2020 was 358 cases, the total number of cigarette sales cases in provinces in 2021 was 358 cases, and each data table contains 15 columns.
[0044] We performed a distribution analysis on the data and deleted the abnormal records in the dataset.
[0045] We used the filtered dataset, and the model training and testing steps are as follows:
[0046] (i) Find the total number of boxes sold and the sales price per box of each cigarette specification in this period. Extract the three columns of "Cigarette Specification", "Month" and "Total" from the sales files from 2011 to 2022, and calculate the total number of boxes sold and the sales price per box of each specification in the current period. Then filter for "Commercial Sales Volume_This Period (Boxes)", if the content of the "Month" column of this row is "Commercial Sales Volume_This Period (Boxes)" and the "Total" is greater than 0, then add up the "Total" to calculate the total number of boxes sold for this cigarette specification; for "Single Box Amount_This Period (Yuan / Box)", if the content of the "Month" column of this row is "Single Box Amount_This Period (Yuan / Box)" and the "Total" is greater than 0, then record the "Total" as the sales price per box.
[0047] (ii) Calculate the total sales amount of each cigarette specification in each province. By reading the sales documents and corresponding cigarette specification price data from 2011 to 2022, filter out the valid sales records of each province and its various cigarette specifications, and then calculate the total sales volume and total sales amount of each cigarette specification (i.e., total sales volume multiplied by the price per box).
[0048] (iii) Calculate the unit price of each cigarette specification based on the total amount. Read the cigarette sales data of each province and calculate the average cigarette unit price of each province from 2011 to 2021. It calculates the average unit price by accumulating the total amount and total number of boxes of all cigarette specifications in each province, and then calculating it by average unit price = total amount / total number of boxes / 250 / 10.
[0049] (iv) Data required to form the model.
[0050] (v) Divide the processed dataset into training set and test set, 80% of the dataset is divided into training set and 20% of the dataset is divided into test set. Build the model.
[0051] (vi) Use the constructed model for training.
[0052] (vii) Save the trained model.
[0053] (viii) Model training effect, such as Figure 2 , Figure 3 and Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , 9 , Figure 10 .
[0054] The mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the error between the predicted value and the true value to evaluate the accuracy of the prediction model. The calculation formula is as follows:
[0055] Calculation formula for Mean Absolute Error (MAE):
[0056] Calculation formula for Root Mean Square Error (RMSE):
[0057] From Figure 2 、 Figure 3 and Figure 4 , Figure 5 、 Figure 6 、 Figure 7 , Figure 8 、 9 、 Figure 10 It can be seen that in the prediction effect of cigarette prices in each province from 2019 to 2021, compared with the prediction effects of the LSTM and LSTM-CNN models, the proposed model of the present invention has better performance in terms of the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) indicators, and the predicted data is more accurate.
[0058] (ix) The model training data is as follows Figure 10 as shown.
[0059] The model training data is the sales data of cigarettes between 2011 and 2018, where the training set accounts for 80% of the total data, and the remaining 20% of the data is used for the test set. The model predicts the cigarette sales prices in 2019, 2020, and 2021 respectively and compares them with the actual prices to determine the model prediction effect. The prediction models used include LSTM, LSTM-CNN, and CRL-Net. The prediction data and prediction effect results of the cigarette prices in each province and city (due to abnormal data in Beijing and Tianjin in 2021, no prediction is made) are as Figure 10 shown.
[0060] (x) The prediction data is as follows Figure 12 as shown.
[0061] Using the LSTM, LSTM-CNN, and CRL-Net models, the cigarette price data results in each province and city (due to abnormal data in Beijing and Tianjin in 2021, no prediction is made) in 2022, 2023, and 2024 are predicted respectively.
[0062] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.
Claims
1. A spatial zoning fast-moving consumer goods placement strategy method based on improved LSTM, characterized in that: include: Obtain historical price data of fast-moving consumer goods, clean the data, and divide it into test sets and training sets; Input the data in the training set into the CRL-Net fast-moving consumer goods price prediction model for data training. The CRL-Net fast-moving consumer goods price prediction model is constructed by combining a convolutional neural network (CNN), a residual connection (ResNet), and a long short-term memory (LSTM) network; Input the data in the test set into the trained CRL-Net fast-moving consumer goods price prediction model, and use the convolutional neural network CNN to extract features from the fast-moving consumer goods historical price data to obtain fast-moving consumer goods price feature data; Using residual connection ResNet to perform data enhancement on the price feature data of the fast-moving consumer goods; Inputting the enhanced price feature data of the fast-moving consumer goods into the long short-term memory LSTM network to dynamically predict the price of the fast-moving consumer goods, and obtaining the price prediction result of the fast-moving consumer goods. The long short-term memory LSTM network is used to capture the long-term dependency in the time series data; The price prediction results of the fast-moving consumer goods are subjected to denormalization processing to obtain the final spatial zoning fast-moving consumer goods launch strategy.
2. The method according to claim 1, characterized in that The method of extracting features from the historical price data of fast-moving consumer goods using a convolutional neural network (CNN) includes: Performing window processing on the historical price data of fast-moving consumer goods to form fast-moving consumer goods price window data; For the fast-moving consumer goods price window data, a convolution operation is used to extract local feature information; The price characteristic data of the fast-moving consumer goods is determined according to the extracted local characteristic information.
3. The method according to claim 1, characterized in that The method of using the residual connection ResNet to perform data enhancement on the price feature data of the fast-moving consumer goods includes: Constructing a plurality of residual blocks, each of the residual blocks comprising two convolutional layers; For the fast-moving consumer goods price feature data, feature transformation is performed through the residual block while retaining the original input features; The transformed features are added to the original input features to obtain a new feature representation.
4. The method according to claim 1, characterized in that The step of inputting the enhanced fast-moving consumer goods price feature data into a long short-term memory (LSTM) network to dynamically predict the price of fast-moving consumer goods includes: Control the degree of forgetting of long-term states through the forget gate; Control the storage level of new input data through the input gate; The output gate is used to control the degree to which the current state is output to the next state; Update the current long-term and short-term states based on the current input and the previous state.
5. The method according to claim 1, characterized in that The denormalization of the fast-moving consumer goods price prediction result includes: Get the maximum and minimum values of normalization; Based on the predicted normalized historical price data of fast-moving consumer goods, combined with the maximum and minimum values, the actual price forecast value of fast-moving consumer goods is calculated; Output the anti-normalized fast-moving consumer goods price prediction results.