Multi-dimensional data statistics intelligent analysis and prediction system and method

Abstract

The invention discloses a multi-dimensional data statistics intelligent analysis and prediction system and method, and particularly relates to the technical field of data prediction. Historical sales data, consumer behavior data, market environment data and external dynamic data of a target analysis object are collected, an LSTM basic prediction model is constructed, consumer behavior change trend features and market environment dynamic features are further extracted, the features are fused with the LSTM model to generate a multi-modal prediction model, and the multi-modal prediction model is analyzed. Comprehensively analyzing the interaction influence of the multi-dimensional data by using a multi-input structure; according to the method, influence weights and contribution degrees of different characteristics are evaluated through sensitivity analysis and interaction analysis, model parameter weights are dynamically adjusted, an optimized prediction model is utilized to generate a multi-dimensional prediction result in a fixed time period, and inventory management and a promotion resource allocation strategy are optimized accordingly, so that long-term drift of consumption behaviors is effectively coped with, and the consumption performance is improved. The inventory cost and the resource waste are reduced, and the operation efficiency and the business decision accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data prediction, and particularly to a multi-dimensional data statistical intelligent analysis and prediction system and method. Background Art

[0002] Multi-dimensional data statistical intelligent analysis and prediction is a technical means that combines big data technology, statistical analysis methods, and artificial intelligence algorithms. It is used to mine valuable patterns and trends from complex multi-dimensional datasets and predict possible future outcomes. By processing large amounts of multi-dimensional data, this technology can help enterprises and research institutions make more accurate decisions in a rapidly changing environment. Multi-dimensional data refers to data with multiple dimensions or features, such as time, location, user attributes, etc. Intelligent analysis and prediction utilize technologies such as machine learning and deep learning to perform pattern recognition and prediction modeling on this data.

[0003] The core of multi-dimensional data statistical intelligent analysis and prediction technology includes three parts: the arrangement of multi-dimensional data structures, the application of statistical analysis methods, and the establishment of intelligent prediction models. First, multi-dimensional data is usually stored in the form of high-dimensional matrices or data cubes, and data cleaning and reduction are required to organize the data to make it suitable for analysis. Then, statistical methods (such as regression analysis, clustering analysis) are used for feature extraction and correlation analysis to find potential patterns. Finally, prediction models are established through intelligent algorithms (such as neural networks, support vector machines, time series analysis) to generate accurate estimates of future development. Taking an online retail platform as an example, it needs to analyze user purchase behavior data (such as purchase time, product type, consumption amount, user geographical location, etc.). By analyzing this multi-dimensional data, the platform can identify user preferences and consumption trends. For example, based on the LSTM model, it can be speculated whether the future sales volume of a certain product will increase significantly during a specific holiday, thereby optimizing inventory management and promotion strategies.

[0004] The existing technologies have the following deficiencies:

[0005] With the evolution of consumers' shopping habits, such as the popularity of online shopping and the decreasing sensitivity to traditional holiday promotions, historical data may not fully reflect these long-term trend changes. If the LSTM model mainly relies on past sales data for prediction, it may assume that future consumer behavior is consistent with historical patterns, thus overestimating demand growth. For example, the sales volume of certain products has increased year by year during past holidays, and the model may extrapolate this trend to the future without considering that consumers are gradually turning to more flexible non-holiday purchases or losing interest in promotional activities. Such misjudgments may lead to enterprises overstocking, occupying inventory costs, and at the same time investing resources in inefficient promotional activities, ultimately resulting in waste and a decline in operational efficiency. Summary of the Invention

[0006] The object of the present invention is to provide a multi - dimensional data statistical intelligent analysis and prediction system and method to solve the deficiencies in the background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: A multi - dimensional data statistical intelligent analysis and prediction method, including the following steps:

[0008] S1: Collect multi - dimensional data of the target analysis object, where the multi - dimensional data includes historical sales data, consumer behavior data, market environment data, and external dynamic data;

[0009] S2: Based on the time - series features in the multi - dimensional data, use the LSTM model to construct a basic prediction model, add holiday feature annotations to the time - series data, and identify the holiday sales pattern of the target object through holiday effect analysis;

[0010] S3: Extract the consumer behavior change trend features and market environment dynamic features from the multi - dimensional data; fuse them with the LSTM model to generate a multi - modal prediction model, and the multi - modal prediction model is based on a multi - input structure to comprehensively analyze the interactive influence of multi - dimensional data;

[0011] S4: Through sensitivity analysis and interaction analysis methods, evaluate the influence weights and contribution degrees of the consumer behavior change trend features and market environment dynamic features on the LSTM model, and according to the evaluation results, dynamically adjust the weights of different parameters in the LSTM model;

[0012] S5: Use the dynamically adapted and optimized prediction model to generate multi - dimensional prediction results of the target analysis object within a fixed time period, and based on the prediction results, optimize the inventory management strategy.

[0013] Preferably, in S2, construct the basic prediction model LSTM. The input layer includes: receiving time - series features and holiday annotations; the LSTM hidden layer includes: multiple stacked LSTM units for extracting deep features of the time series; the fully - connected layer: reducing the dimension of the high - dimensional features output by the hidden layer and generating prediction values; the output layer: predicting future sales volume values; use historical data to train the model, optimize the objective function, evaluate the model performance through cross - validation, and adjust hyperparameters; introduce holiday feature annotations for the time - series data to explicitly represent the influence of holidays on sales volume. The holiday features are used as additional input dimensions and jointly input into the LSTM model with the time - series data, and the holiday features are transformed into a trainable vector representation.

[0014] Preferably, in S3, after analyzing the search - heat change rate in the extracted consumer behavior change trend features, generate a search - heat fluctuation index. The acquisition method of the search - heat fluctuation index is:

[0015] The input is the time - series data of the search heat, denoted as: ; where, represents the search popularity on the t-th day, H is the total time length, the size w of the sliding window is set, and for each time point t, a window is defined, including data from the (t - w + 1)-th day to the t-th day: ; Calculate the standard deviation of the search popularity on the window and perform normalization: ; ; and are respectively the minimum and maximum values of the standard deviations of all windows, is the normalized standard deviation value, and the normalized standard deviation is used as the search popularity fluctuation index.

[0016] Preferably, in S3, after analyzing the abnormal situation of the competitor price fluctuation in the dynamically extracted market environment features, a competitor price anomaly index is generated. The method for obtaining the competitor price anomaly index is as follows:

[0017] Set the competitor price time series as: ; where, is the competitor price at the Q-th time point, Q is the total length of the time series, and calculate the price change rate as the input feature for detection: ; is the price change rate on the t-th day; the data distribution model represents the price change rate, where the boundary of each paragraph is defined as a change point, and the prior distribution of the change point is set, and the expression is: ; is the average interval time for the change point to occur, is the prior distribution, and it is set that the price change rate follows a normal distribution within each paragraph, and the expression is: ; is the mean value of the price change rate in each paragraph, is the variance of the price change rate in each paragraph, and use Bayes' formula to calculate the posterior probability that each time point is a change point , and the expression is: ; In the formula, is the price change rate sequence, represents the likelihood of the sequence under the assumption that t is a change point, and the calculated posterior probability is used as the competitor price anomaly index.

[0018] Preferably, in S3, the search popularity fluctuation index and the competitor price anomaly index are fused with the LSTM model to generate a multi-modal prediction model. The multi-modal prediction model is based on a multi-input structure to comprehensively analyze the interactive effects of multi-dimensional data, including:

[0019] Taking the search popularity fluctuation index, the competitor price anomaly index, and the historical sales data of the target product in the LSTM model as input items, a multi-input structure is constructed. The inputs of different modalities are processed separately, and each input enters a different LSTM sub-module for feature extraction;

[0020] Use the LSTM layer to process the sales volume time series Y and output the sales volume feature representation : ;

[0021] Use the LSTM layer to process the search popularity fluctuation index H and output the search popularity feature representation : ;

[0022] Use the LSTM layer to process the competitor price anomaly index A and output the price anomaly feature representation : ;

[0023] Fuse the features extracted by different sub-modules to construct a global feature representation ; Combine into a unified feature vector: ;

[0024] Pass the fused features into the fully connected layer to further extract high-dimensional non-linear features;

[0025] Output layer: Predict the future sales volume value, and the expression is: ;

[0026] Use the mean squared error as the loss function to measure the deviation between the predicted value and the actual value.

[0027] Preferably, in S4, a small perturbation is made to the search popularity fluctuation index, and the perturbed features are input into the LSTM model, and the model output value is recorded ; Calculate the prediction difference before and after the perturbation: ; For each feature, calculate its average sensitivity to the prediction result , and the expression is: ; The larger the value, the greater the influence of the feature on the prediction result. N is the total number of sample points in the time series, that is, the total number of time points in the input data. Normalize the sensitivities of all features to generate the feature importance weights , and the expression is: ; M is the total number of features;

[0028] Interaction analysis evaluates the combined impact between different features. By perturbing two features simultaneously, including the search popularity fluctuation index and the competitor price anomaly index, the changes in model predictions are observed: ; represents the feature the prediction change caused by the combined perturbation, is the model output value of the feature . Calculate the interaction contribution of the feature : ; In the formula, represents the contribution degree that the synergistic impact of exceeds the individual impact. Normalize the interaction contributions of all feature pairs to generate the interaction weight .

[0029] Preferably, according to the sensitivity weight and the interaction weight , dynamically adjust the feature fusion weight in the LSTM model. The update formula for the fusion weight: ; is the final weight of the k-th feature, η∈[0,1] is the importance ratio for adjusting the sensitivity and the interaction weight; Apply the weight to the multi-modal feature fusion layer: ; where is the extraction result of each modal feature.

[0030] Preferably, in S5, use the dynamically adapted and optimized prediction model f to generate the multi-dimensional prediction results of the target analysis object within a fixed time period. The formula is: ; In the formula, represents the target prediction result from time t to t+T, represents the sales volume prediction within the next T time units, f is the optimized multi-modal prediction model, represents the sales volume time series feature at time t, represents the search popularity fluctuation index at time t, represents the competitor price anomaly index at time t, Θ represents the set of model parameters. According to the prediction result , optimize the inventory management, calculate the total demand within the time period from t to t+T. The expression is: ; In the formula, is the predicted total demand, is the predicted sales volume at time i, T is the length of the prediction time range;

[0031] Calculate the safety stock S: : Z is the service level factor, σ is the standard deviation of the prediction error, calculate the target inventory level within the time period , the expression is: ; represents the target inventory level within the prediction time period, which is used to guide the replenishment decision-making; if the current inventory is lower than the target inventory , then the replenishment quantity is: ; where R is the replenishment quantity, is the current inventory.

[0032] The present invention also provides a multi-dimensional data statistical intelligent analysis and prediction system, including a data collection module, a time series prediction module, a multi-modal fusion module, a dynamic optimization module, and a prediction and management module:

[0033] Data collection module: Collect multi-dimensional data of the target analysis object, and the multi-dimensional data includes historical sales data, consumer behavior data, market environment data, and external dynamic data;

[0034] Time series prediction module: Based on the time series characteristics in the multi-dimensional data, use the LSTM model to construct a basic prediction model, add holiday feature annotations to the time series data, and identify the holiday sales pattern of the target object through holiday effect analysis;

[0035] Multi-modal fusion module: Extract the consumer behavior change trend characteristics and market environment dynamic characteristics from the multi-dimensional data; fuse them with the LSTM model to generate a multi-modal prediction model, and the multi-modal prediction model is based on a multi-input structure to comprehensively analyze the interaction effects of multi-dimensional data;

[0036] Dynamic optimization module: Through sensitivity analysis and interaction analysis methods, evaluate the influence weights and contribution degrees of the consumer behavior change trend characteristics and market environment dynamic characteristics on the LSTM model, and dynamically adjust the weights of different parameters in the LSTM model according to the evaluation results;

[0037] Prediction and management module: Use the dynamically adapted and optimized prediction model to generate multi-dimensional prediction results of the target analysis object within a fixed time period, and optimize the inventory management strategy based on the prediction results.

[0038] In the above technical solution, the technical effects and advantages provided by the present invention:

[0039] 1. The present invention collects historical sales data, consumer behavior data, market environment data, and external dynamic data, constructs a basic prediction model by combining with the LSTM model, and introduces holiday feature annotation and holiday effect analysis, which can accurately identify the impact mode of holidays on sales volume. In addition, by extracting the characteristics of the changing trend of consumer behavior (such as the search heat fluctuation index) and the dynamic characteristics of the market environment (such as the abnormal price index of competing products), and fusing them with the LSTM model to generate a multi-modal prediction model, comprehensively analyzing the interactive effects of multi-dimensional data, effectively improving the accuracy and robustness of the prediction.

[0040] 2. The present invention uses sensitivity analysis and interaction analysis methods to evaluate the contribution degree of different features to the prediction results, and dynamically adjusts the feature fusion weights in the model, enabling the model to have the ability to dynamically adapt to changes in consumer behavior and the market environment. Finally, using the optimized multi-modal prediction model to generate multi-dimensional prediction results within a fixed time period, and accordingly optimizing the inventory management strategy (such as calculating the total demand, safety stock, and replenishment quantity), reducing the risk of overstock or shortage; at the same time, dynamically adjusting the allocation of promotion resources to improve the utilization efficiency of resources. The present invention significantly improves the enterprise operation efficiency and market response ability while reducing inventory costs and resource waste, providing strong support for intelligent decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to these drawings.

[0042] Figure 1 It is a flowchart of the method of the present invention.

[0043] Figure 2 It is a system module diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0045] Example 1, please refer to Figure 1 As shown, the multi-dimensional data statistical intelligent analysis and prediction method in this embodiment includes the following steps:

[0046] S1: Collect multi-dimensional data of the target analysis object, where the multi-dimensional data includes historical sales data, consumer behavior data, market environment data, and external dynamic data;

[0047] S2: Based on the time series characteristics in the multi-dimensional data, use the LSTM model to construct a basic prediction model, add holiday feature annotations to the time series data, and identify the holiday sales patterns of the target object through holiday effect analysis;

[0048] S3: Extract the consumer behavior change trend characteristics and market environment dynamic characteristics from the multi-dimensional data; fuse them with the LSTM model to generate a multi-modal prediction model, and the multi-modal prediction model is based on a multi-input structure to comprehensively analyze the interactive effects of multi-dimensional data;

[0049] S4: Through sensitivity analysis and interaction analysis methods, evaluate the influence weights and contribution degrees of the consumer behavior change trend characteristics and market environment dynamic characteristics on the LSTM model, and according to the evaluation results, dynamically adjust the weights of different parameters in the LSTM model;

[0050] S5: Use the dynamically adapted and optimized prediction model to generate multi-dimensional prediction results of the target analysis object within a fixed time period, and based on the prediction results, optimize the inventory management strategy.

[0051] In S1, the historical sales data is the sales situation of the target product within a specific time period, usually presented in the form of a time series, including: Time dimension: sales date, time (such as daily, weekly, monthly sales volume). Product information: product category, SKU (stock keeping unit), price range. Sales channels: online sales platforms, offline stores, third-party distributors. Promotion information: whether it is in the promotion period, promotion methods (discounts, full reduction, gifts, etc.). The data source is: the enterprise internal ERP system or the sales record database. Capture historical sales trends, periodic patterns, and abnormal peaks as the basic input for the prediction model.

[0052] Consumer behavior data is dynamic data used to reflect the interaction behavior between consumers and the target product, reflecting the changes in consumer needs and interests. The main contents include: Browsing behavior: the number of clicks on the product page, the stay duration, and the bounce rate. Search behavior: the search volume of relevant keywords and the time series change of search popularity. Purchase behavior: purchase frequency, repurchase rate, average customer unit price. Evaluation and feedback: the score of the product evaluation, the text content (extract positive or negative emotions through sentiment analysis). The data sources include: e-commerce platform data (such as click stream logs, search logs). Third-party consumer research reports.

[0053] Market environment data is used to describe the external environmental variables of the market where the target product is located, including industry dynamics and changes in the competitive landscape. The main contents include: Industry competition data: market share, pricing strategies of major competitors, new product launch dynamics. Price fluctuation data: price change trends of similar products in the market and the impact of promotional activities. Market capacity: consumer scale, growth or contraction of the target market. Policies and regulations: changes in laws and policies affecting sales (such as tax adjustments, import restrictions). Data sources include: industry reports (such as Nielsen, iResearch, etc.). Public data sets (such as government statistical bureaus or trade associations).

[0054] External dynamic data is for non-direct factors that may be related to the sales volume of the target product, including macroeconomic indicators and social dynamics. The main contents include: Macroeconomic data: GDP growth rate, consumer confidence index, inflation rate. Weather data: temperature, precipitation, seasonal changes (especially for sensitive products such as clothing and beverages). Social hotspots: fashion trends, social media discussion heat (extracted through keywords). Event impacts: special events or disasters (such as the impact of the epidemic, natural disasters on consumer behavior). Data sources include: economic reports released by the government, weather data from meteorological departments. Data mining tools on social media platforms (such as Twitter, Weibo).

[0055] In this application, the collection of multi-dimensional data not only includes direct sales data, but also comprehensive information on consumer behavior, market environment, and external dynamic data. These data together constitute the input basis for model prediction. By comprehensively integrating these data sources, the interaction between multi-dimensional variables can be captured, providing support for more accurate analysis and prediction.

[0056] In S2, time series features are extracted from the multi-dimensional data, such as daily sales volume, price changes, inventory levels, etc. Decompose the time series data: Trend component: reflecting the long-term change trend of sales volume. Seasonal component: capturing the periodic fluctuations of sales volume, such as the periodic impact of holidays on sales volume. Residual component: containing the impact of random fluctuations or unexpected events.

[0057] Data preprocessing includes: Data cleaning: handling missing values, outliers (such as abnormal sales volume fluctuations during holidays). Normalization processing: normalizing the time series data to a unified numerical range for convenient LSTM model training. Time window division: using the sliding window technique to generate training samples, taking the data of the past n days as input and the data of the future m days as output.

[0058] Build a basic prediction model (LSTM). LSTM (Long Short-Term Memory network) is suitable for processing time series data and has the following advantages: It can capture dependencies over long time spans (such as the long-term impact of holidays on sales). It can handle non-linear and highly noisy data and adapt to complex time series patterns.

[0059] Input layer: Receive time series features (such as historical sales, prices) and holiday annotations (see Step 3). LSTM hidden layer: Stack multiple LSTM cells to extract deep features of the time series. Fully connected layer: Reduce the dimensionality of the high-dimensional features output by the hidden layer and generate prediction values. Output layer: Predict future sales values (such as daily or weekly sales).

[0060] Train the model using historical data and optimize the objective function (such as Mean Squared Error MSE). Evaluate the model performance through cross-validation and adjust hyperparameters (such as the number of hidden layer units, learning rate).

[0061] Introduce holiday feature annotations for time series data to explicitly represent the impact of holidays on sales: Holiday type: Mark whether each day is a holiday (such as "1" represents a holiday, "0" represents a non-holiday). Number of days before and after the holiday: Introduce features to represent the number of days from the holiday (such as 3 days before the holiday, 5 days after the holiday). Holiday classification: Distinguish the types of holidays (such as legal holidays, shopping festivals, promotional event days).

[0062] Use the holiday features as an additional input dimension and jointly input them into the LSTM model with the time series data. Use the Embedding technique to transform the holiday features into a trainable vector representation to enhance the model's learning ability for holiday patterns.

[0063] Utilize statistical analysis and the model prediction results to identify the significant impact of holidays on sales: Analysis of sales volume changes before and after holidays: Compare the sales volume changes before, during, and after the holiday and calculate the increment amplitude. Statistical test: Use t-test or Analysis of Variance (ANOVA) to verify the significance of the holiday effect.

[0064] Based on the holiday effect analysis, identify the typical impact patterns of different holidays on sales: Peak effect: Such as "Double Eleven", the sales volume surges significantly on the holiday day. Early warm-up effect: Such as Christmas, the sales volume gradually increases before the holiday. Follow-up afterglow effect: Such as the Spring Festival, the sales volume remains relatively high after the holiday ends.

[0065] Use the results of the holiday effect analysis as an important reference for the prediction model, adjust the prediction weights during holidays (such as increasing the sales volume expectation for peak holidays). Post-process the model prediction results and adjust the inventory replenishment and promotional activity design according to the holiday type.

[0066] In this application, through the extraction of time series features and the construction of an LSTM model, combined with holiday feature annotation and holiday effect analysis, the holiday sales pattern of the target object can be accurately identified. This method not only captures the complex dependencies in the time series but also comprehensively considers the unique impact of holidays on sales, providing technical support for optimizing prediction accuracy and business decisions.

[0067] In S3, extract the consumer behavior change trend features and market environment dynamic features from the multi-dimensional data, including:

[0068] The consumer behavior change trend features reflect the dynamic changes in consumer demand preferences and are an important basis for predicting future demand. Search behavior features: Keyword search volume: The number of times consumers search for relevant keywords on search engines or platforms. Search heat change rate: The increase or decrease in search heat compared to the previous period. Purchase behavior features: Repurchase rate: The proportion of users who have purchased the product who repurchase it within a specific time. Shopping frequency: The average number of purchases made by users within a certain time. Interest features: Click-through rate and conversion rate: The number of clicks on the product page and the purchase proportion after clicking. Stay duration: The average stay time of users on the product page, reflecting the depth of interest. Evaluation features: Sentiment analysis score: Conduct sentiment analysis on user evaluation texts to extract sentiment tendencies (such as positive, neutral, negative). Evaluation keyword extraction: Identify high-frequency keywords to reflect consumer concerns. Data sources: E-commerce platform logs (such as search, click, and purchase records). Comment and interaction data from social media (such as Twitter, Weibo). Tools and technologies: Use natural language processing (NLP) technologies to process text data (such as sentiment analysis, keyword extraction). Analyze behavioral log data to extract indicators such as clicks, stays, and searches.

[0069] Feature processing includes: Time serialization: Convert data such as search volume and evaluation changes into time series to reflect dynamic change trends. Normalization processing: Normalize indicators such as frequency and score for analysis under the same dimension. Feature aggregation: Calculate the mean or change rate according to time windows (such as weeks, months) to simplify the analysis. The market environment dynamic features reveal the impact of the external environment on commodity demand and are usually macroscopic and competitive.

[0070] After analyzing the search heat change rate in the extracted consumer behavior change trend features, a search heat fluctuation index is generated. The method for obtaining the search heat fluctuation index is as follows:

[0071] The input is the time series data of search heat, which represents the change in search heat within a specific time period, denoted as: ; where Denote the search popularity on the t-th day, and H is the total time length. Set the size of the sliding window as w, which represents the number of days covered for calculating the standard deviation each time (e.g., w = 7 represents one week).

[0072] For each time point t, define a window that contains the data from the (t - w + 1)-th day to the t-th day: ; where when t < w, the window cannot be calculated and needs to be filled as appropriate. Calculate the standard deviation of the search popularity on the window ; To map the standard deviation values at different time points to a unified range (e.g., [0, 1]), perform normalization: ; ; and are the minimum and maximum values of the standard deviations of all windows respectively, is the standard deviation value after normalization, and use the normalized standard deviation as the search popularity fluctuation index.

[0073] The categories of the dynamic characteristics of the market environment include: Competitive intensity characteristics: Number of competitors: The number of brands or products directly competing with the target product. Price fluctuations of competing products: The price change trend of similar products, especially the discount intensity during promotional periods. Market trend characteristics: Change in market share: The change in the proportion of the target product or brand in the market. New product release dynamics: The release time and features of new products similar to the target product. Economic dynamic characteristics: Consumer confidence index: Reflects consumers' expectations for the future economic situation. Per capita disposable income: The impact of the economic environment on consumers' purchasing power. External event characteristics: Weather changes: Such as the impact of temperature on the demand for cold drinks or warm clothing. Social hot events: Such as the driving effect of festivals or emergencies on purchasing behavior. Data sources: Industry reports (such as Nielsen, iResearch). Macroeconomic databases (such as government statistical bureaus, international economic databases). Meteorological data platforms or social media event monitoring tools. Tools and technologies: Use web crawler technology to collect market dynamic data (such as prices of competing products, consumer public opinions). Use APIs to obtain meteorological and economic indicator data.

[0074] Feature processing includes: Time alignment: Serialize the market environment data in time to ensure it matches the sales data. Classification processing: Label qualitative data (such as social events) with event category tags for easy model processing. Weighted analysis: Assign weights to different market environment characteristics to quantify their impact on demand.

[0075] After analyzing the abnormal situation of price fluctuations of competing products in the extracted dynamic characteristics of the market environment, generate a competing product price anomaly index. The method for obtaining the competing product price anomaly index is:

[0076] Set the time series of competing product prices as: ; where is the price of competing products at the Qth time point, where Q is the total length of the time series. Calculate the price change rate as the main input feature for detection: ; is the price change rate on the tth day. It is assumed that the distribution of the price change rate is different in different time periods but remains consistent within the same period. Use a piecewise constant distribution model to represent the price change rate, where the boundary of each segment is defined as a change point. Set the prior distribution of the change point, and the expression is: ; is the average interval time when the change point occurs, is the prior distribution. Assume that the price change rate follows a normal distribution within each segment, and the expression is: ; is the mean of the price change rate for each segment, is the variance of the price change rate for each segment. Use Bayes' formula to calculate the posterior probability that each time point is a change point , and the expression is: ; In the formula, is the price change rate sequence, represents the likelihood of the sequence under the assumption that t is a change point. Take the calculated posterior probability as the competing product price anomaly index.

[0077] Fuse the search popularity fluctuation index and the competing product price anomaly index with the LSTM model to generate a multi-modal prediction model. The multi-modal prediction model is based on a multi-input structure to comprehensively analyze the interactive effects of multi-dimensional data, including:

[0078] Take the search popularity fluctuation index, the competing product price anomaly index, and the historical sales data of the target product in the LSTM model as input items to construct a multi-input structure, and process the inputs of different modalities separately: Input 1: The time series Y of the target product sales volume. Input 2: The time series H of the search popularity fluctuation index. Input 3: The time series A of the competing product price anomaly index. Each input enters a different LSTM sub-module for feature extraction.

[0079] Use the LSTM layer to process the sales volume time series Y and extract its long-term trend and periodic characteristics.

[0080] Output: Sales volume feature representation : ;

[0081] Search popularity fluctuation feature extraction: Use the LSTM layer to process the search popularity fluctuation index H and extract the dynamic characteristics of consumer behavior.

[0082] Output: Search popularity feature representation : ;

[0083] Abnormal Feature Extraction of Competitor Prices: Use the LSTM layer to process the abnormal index A of competitor prices and extract the abnormal features of the market environment.

[0084] Output: Price Abnormal Feature Representation : .

[0085] Fuse the features extracted by different sub-modules to construct a global feature representation ; Concatenate into a unified feature vector: ;

[0086] Pass the fused features into the fully connected layer (Dense Layer) to further extract high-dimensional non-linear features.

[0087] Output Layer: Predict the future sales volume value, and the expression is: .

[0088] Use the mean squared error (MSE) as the loss function to measure the deviation between the predicted value and the actual value; select an optimizer suitable for time series tasks (such as Adam) for parameter optimization.

[0089] Data Partitioning: Divide the data into training set, validation set and test set (such as 7:2:1).

[0090] Batch Training: Use small batch data (Batch) for iterative training to improve the convergence efficiency.

[0091] Hyperparameter Tuning: Adjust parameters such as learning rate, number of LSTM hidden units, and time window size.

[0092] Use the following metrics to evaluate the model performance: Mean Squared Error (MSE): Measure the overall error between the predicted value and the true value.

[0093] Mean Absolute Percentage Error (MAPE): Measure the relative size of the prediction error.

[0094] Coefficient of Determination (R²): Evaluate the goodness of fit of the model to the data.

[0095] Verify the contribution of different modality features to the prediction: Remove a certain modality feature and observe the performance change (ablation experiment). Use SHAP values to analyze the importance of features.

[0096] In S4, sensitivity analysis evaluates the response degree of the model output (such as predicted sales volume) to the change of input features, and is used to quantify the impact of each feature on the model performance.

[0097] Make a small perturbation to a certain feature (such as the search popularity fluctuation index), while keeping other features unchanged.

[0098] The perturbation range is set to [μ−kσ,μ+kσ], where μ is the mean of the feature, σ is the standard deviation, and k is an adjustment factor (such as 0.1 or 0.2).

[0099] Input the perturbed feature into the LSTM model and record the model output value ; Calculate the prediction difference before and after the perturbation: ; For each feature, calculate its average sensitivity to the prediction result , and the expression is: ; The larger the value, the greater the impact of the feature on the prediction result. N is the total number of sample points in the time series, that is, the total number of time points in the input data. Normalize the sensitivities of all features to generate feature importance weights , and the expression is: ; M is the total number of features.

[0100] Interaction analysis evaluates the joint impact between different features, that is, the contribution of the co-variation of two or more features to the model output. Perturb two features simultaneously, including the search popularity fluctuation index and the abnormal competitor price index, and observe the model prediction changes: ; represents the feature The prediction change caused by the joint perturbation, is the model output value of the feature , calculate the interaction contribution of the feature : ; In the formula, represents The contribution degree of the co-influence of , and the expression is: .

[0101] According to the sensitivity weight and the interaction weight , dynamically adjust the feature fusion weight in the LSTM model. The update formula of the fusion weight: ; is the final weight of the k-th feature, and η∈[0,1] is the importance ratio for adjusting the sensitivity and interaction weight.

[0102] Apply the weight to the multi-modal feature fusion layer: ; Among them, is the extraction result of each modal feature.

[0103] Retrain the model using the dynamically adjusted weights and observe the changes in model performance. Compare the model performance (such as MSE, MAPE) before and after dynamic weight adjustment. Analyze the enhancement effect on key features after adjustment.

[0104] Feature sensitivity weights and interaction weights , the fused weights after dynamic adjustment . Verify the performance improvement of the optimized LSTM model prediction results in the actual business scenario.

[0105] In S5, use the dynamically adapted and optimized prediction model f to generate multi-dimensional prediction results of the target analysis object within a fixed time period. The formula is: ; where represents the target prediction result from time t to t+T, indicating the sales volume prediction within the next T time units. f is the optimized multi-modal prediction model that fuses different data features (time series, search popularity fluctuation index, competitor price anomaly index). represents the sales volume time series feature at time t, indicating the historical sales trend and seasonal features, represents the search popularity fluctuation index at time t, reflecting the changes in consumer behavior, represents the competitor price anomaly index at time t, reflecting the dynamic features of the market environment. Θ represents the set of model parameters, including feature fusion weights, LSTM network parameters, etc.

[0106] Based on the prediction results , optimize inventory management to ensure supply-demand balance and reduce inventory costs at the same time. Calculate the total demand within the time period from t to t+T , and the expression is: ; where is the predicted total demand, used to guide inventory stocking, is the predicted sales volume at time i, and T is the length of the prediction time range.

[0107] To cope with prediction errors and sudden demands, calculate the safety stock S: : Z is the service level factor, related to the target service level of the supply chain (e.g., for a 95% service level, Z≈1.65). σ is the standard deviation of the prediction error, calculated based on historical data.

[0108] Calculate the target inventory level within the time period , and the expression is: ; represents the target inventory level within the prediction time period, used to guide replenishment decisions. If the current inventory level is lower than the target inventory , then the replenishment quantity is: ; where R is the replenishment quantity, and is the current inventory level.

[0109] In this embodiment, by collecting multi-dimensional data of the target analysis object (including historical sales data, consumer behavior data, market environment data, and external dynamic data), an LSTM-based basic prediction model based on time series features is constructed, and combined with holiday feature annotation and holiday effect analysis, the holiday sales pattern of the target object is identified. On this basis, the characteristics of the change trend of consumer behavior and the dynamic characteristics of the market environment are extracted, and fused with the LSTM model to generate a multi-modal prediction model, which comprehensively analyzes the interactive effects of multi-dimensional data with a multi-input structure. Through sensitivity analysis and interaction analysis methods, the influence weights and contribution degrees of different features on the model are evaluated, and the weights of the model parameters are dynamically adjusted to optimize the prediction accuracy. Finally, the optimized prediction model is used to generate multi-dimensional prediction results within a fixed time period, and based on this, the inventory management strategy is optimized to achieve intelligent resource allocation and decision support.

[0110] Example 2, please refer to Figure 2 As shown, the multi-dimensional data statistical intelligent analysis and prediction system in this embodiment includes a data collection module, a time series prediction module, a multi-modal fusion module, a dynamic optimization module, and a prediction and management module;

[0111] Data collection module: Collect multi-dimensional data of the target analysis object, and the multi-dimensional data includes historical sales data, consumer behavior data, market environment data, and external dynamic data;

[0112] Time series prediction module: Based on the time series features in the multi-dimensional data, an LSTM model is used to construct a basic prediction model, holiday feature annotation is added to the time series data, and the holiday sales pattern of the target object is identified through holiday effect analysis;

[0113] Multi-modal fusion module: Extract the characteristics of the change trend of consumer behavior and the dynamic characteristics of the market environment from the multi-dimensional data; fuse them with the LSTM model to generate a multi-modal prediction model, and the multi-modal prediction model comprehensively analyzes the interactive effects of multi-dimensional data based on a multi-input structure;

[0114] Dynamic optimization module: Through sensitivity analysis and interaction analysis methods, evaluate the influence weights and contribution degrees of the characteristics of the change trend of consumer behavior and the dynamic characteristics of the market environment on the LSTM model, and according to the evaluation results, dynamically adjust the weights of different parameters in the LSTM model;

[0115] Prediction and management module: Use the dynamically adapted and optimized prediction model to generate multi-dimensional prediction results of the target analysis object within a fixed time period, and optimize the inventory management strategy based on the prediction results.

[0116] As described above, this is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application.

Claims

1. A multidimensional data statistical intelligent analysis and prediction method, characterized by: The following steps are involved: S1: Collect multidimensional data of the target analysis object, wherein the multidimensional data includes historical sales data, consumer behavior data, market environment data and external dynamic data; S2: Based on the time series features in multidimensional data, the LSTM model is used to build a basic prediction model, holiday feature annotations are added to the time series data, and the holiday sales pattern of the target object is identified through holiday effect analysis; S3: Extract the changing trend characteristics of consumer behavior and the dynamic characteristics of the market environment from multidimensional data; It is integrated with the LSTM model to generate a multimodal prediction model, which is based on a multi-input structure to comprehensively analyze the interactive impact of multidimensional data; S4: Through sensitivity analysis and interactive analysis methods, evaluate the influence weight and contribution of the consumer behavior change trend characteristics and market environment dynamic characteristics on the LSTM model, and dynamically adjust the weights of different parameters in the LSTM model according to the evaluation results; S5: Use the dynamically adaptive optimized forecasting model to generate multi-dimensional forecasting results for the target analysis object within a fixed time period, and optimize the inventory management strategy based on the forecasting results.

2. The multidimensional data statistical intelligent analysis and prediction method according to claim 1, characterized in that: In S2, a basic prediction model LSTM is constructed. The input layer includes: receiving time series features and holiday labels; the LSTM hidden layer includes: multiple layers of stacked LSTM units, which are used to extract deep features of the time series; the fully connected layer: reduces the dimension of the high-dimensional features output by the hidden layer and generates predicted values; the output layer: predicts future sales values; uses historical data to train the model, optimizes the objective function, evaluates the model performance through cross-validation, and adjusts hyperparameters; introduces holiday feature labels for time series data to explicitly indicate the impact of holidays on sales. Holiday features are used as additional input dimensions and are jointly input into the LSTM model with time series data to convert holiday features into trainable vector representations.

3. The multidimensional data statistical intelligent analysis and prediction method according to claim 2 is characterized by: In S3, the search heat fluctuation index is generated after analyzing the search heat change rate in the extracted consumer behavior change trend characteristics. The method for obtaining the search heat fluctuation index is: The input is the time series data of search popularity recorded as: ;in, represents the search popularity on the tth day, H is the total time length, and the size of the sliding window is set to w. For each time point t, a window is defined, containing data from the t-w+1th day to the tth day: ; In the window Calculate the standard deviation of search popularity , and normalize it: ; and are the minimum and maximum standard deviations of all windows, respectively. is the standard deviation value after normalization, and the normalized standard deviation is used as the search heat fluctuation index.

4. The multidimensional data statistical intelligent analysis and prediction method according to claim 3 is characterized by: In S3, the abnormal price fluctuation of competing products in the extracted dynamic characteristics of the market environment is analyzed to generate a competitive product price abnormality index. The method for obtaining the competitive product price abnormality index is as follows: Set the competitor price time series to: ;in, is the price of the competitor product at the Qth time point, Q is the total length of the time series, and the price change rate is calculated as the input feature of the detection: ; is the price change rate on the tth day; the data distribution model represents the price change rate, where the boundary of each paragraph is defined as a change point, and the prior distribution of the change point is set, and the expression is: ; is the average interval between change points, As the prior distribution, the price change rate is assumed to follow a normal distribution in each segment, and the expression is: ; is the mean price change rate of each period, The variance of the price change rate for each period is used to calculate the posterior probability of each time point being a change point using the Bayesian formula. , the expression is: ; In the formula, is the price change rate series, It represents the likelihood of the sequence under the assumption that t is the change point, and the calculated posterior probability is used as the abnormal price index of the competing product.

5. The multidimensional data statistical intelligent analysis and prediction method according to claim 4 is characterized in that: In S3, the search popularity fluctuation index and the competitor price anomaly index are integrated with the LSTM model to generate a multimodal prediction model. The multimodal prediction model is based on a multi-input structure to comprehensively analyze the interactive impact of multidimensional data, including: The search popularity fluctuation index, competitor price anomaly index, and historical sales data of the target product in the LSTM model are used as input items to build a multi-input structure, process inputs of different modes separately, and each input enters a different LSTM submodule for feature extraction; Use the LSTM layer to process the sales time series Y and output the sales feature representation : ; Use the LSTM layer to process the search heat fluctuation index H and output the search heat feature representation : ; Use the LSTM layer to process the competitor price anomaly index A and output the price anomaly feature representation : ; The features extracted by different sub-modules are integrated to construct a global feature representation. ;Will Concatenate into a unified feature vector: ; The fused features Passed into the fully connected layer to further extract high-dimensional nonlinear features; Output layer: predict future sales The value of is expressed as: ; The mean square error is used as the loss function to measure the deviation between the predicted value and the actual value.

6. The multidimensional data statistical intelligent analysis and prediction method according to claim 5, characterized in that: In S4, a small range perturbation is performed on the search heat fluctuation index, the perturbed features are input into the LSTM model, and the model output value is recorded. ; Calculate the difference in predictions before and after the perturbation: ; For each feature, calculate its average sensitivity to the prediction result , the expression is: ; The larger the value, the greater the impact of the feature on the prediction result. N is the total number of sample points in the time series, that is, the total number of time points in the input data. The sensitivity of all features is normalized to generate feature importance weights. , the expression is: ; M is the total number of features; Interactive analysis evaluates the joint impact of different features. Two features are disturbed at the same time, including the search popularity fluctuation index and the competitor price anomaly index, to observe the changes in model predictions: ; Representation characteristics The change in forecast caused by the joint disturbance, Features The model output value of , calculate the feature Interactive contributions: ; In the formula, express The synergistic effect exceeds the contribution of the individual effect, and the interaction contribution of all feature pairs is normalized to generate the interaction weight , the expression is: .

7. The multidimensional data statistical intelligent analysis and prediction method according to claim 6, characterized in that: Based on sensitivity weight and interaction weights , dynamically adjust the feature fusion weights in the LSTM model, and the update formula of the fusion weights is: ; is the final weight of the kth feature, η∈[0,1] is the importance ratio of adjustment sensitivity and interaction weight; Applied to multimodal feature fusion layer: ;in The extraction results of each modal feature.

8. The multidimensional data statistical intelligent analysis and prediction method according to claim 7, characterized in that: In S5, the prediction model f after dynamic adaptive optimization is used to generate the multi-dimensional prediction results of the target analysis object within a fixed time period. The formula is: ; In the formula, represents the target prediction result in the time period t to +T, represents the sales forecast in the next T time units, f is the optimized multimodal prediction model, represents the sales time series characteristics at time t, Represents the search popularity fluctuation index at time t, represents the abnormal price index of competing products at time t, Θ represents the parameter set of the model, and according to the prediction results , optimize inventory management and calculate the total demand from time period t to t+T , the expression is: ; In the formula, is the total predicted demand, is the predicted sales volume at time i, T is the length of the prediction time range; Calculate the safety stock S: :Z is the service level factor, σ is the standard deviation of the forecast error, and the target inventory level within the time period is calculated , the expression is: ; Indicates the target inventory level during the forecast period, which is used to guide replenishment decisions; if the current inventory Below target inventory , then the replenishment quantity is: ; In the formula, R is the replenishment quantity, The current inventory quantity.

9. A multidimensional data statistical intelligent analysis and prediction system, used to implement the multidimensional data statistical intelligent analysis and prediction method according to any one of claims 1 to 8, characterized in that: Including data collection module, time series prediction module, multimodal fusion module, dynamic optimization module and prediction and management module: Data collection module: collects multi-dimensional data of the target analysis object, including historical sales data, consumer behavior data, market environment data and external dynamic data; Time series prediction module: Based on the time series features in multidimensional data, the LSTM model is used to build a basic prediction model, holiday feature annotations are added to the time series data, and the holiday sales pattern of the target object is identified through holiday effect analysis; Multimodal fusion module: extracts the changing trend characteristics of consumer behavior and the dynamic characteristics of the market environment from multidimensional data; It is integrated with the LSTM model to generate a multimodal prediction model, which is based on a multi-input structure to comprehensively analyze the interactive impact of multidimensional data; Dynamic optimization module: through sensitivity analysis and interactive analysis methods, evaluate the influence weight and contribution of the trend characteristics of consumer behavior changes and the dynamic characteristics of the market environment on the LSTM model, and dynamically adjust the weights of different parameters in the LSTM model according to the evaluation results; Forecasting and management module: Utilize the dynamically adaptive optimized forecasting model to generate multi-dimensional forecasting results for the target analysis object within a fixed time period, and optimize the inventory management strategy based on the forecasting results.

Images