A multi-model adaptive fusion flower price index prediction method and system

Abstract

This invention discloses a multi-model adaptive fusion method and system for predicting flower price indices, belonging to the fields of smart agriculture and flower big data technology. Addressing the strong seasonality, holiday effects, significant varietal differences, and lag in price transmission characteristic of flowers, this invention proposes specific data preprocessing: distinguishing outliers from holiday fluctuations and supplementing missing values ​​with industry experience; constructing derived features such as demand intensity and supply risk; designing scenario-driven dynamic model routing, configuring models and factor weights differently according to short / medium / long-term periods, with long-term prediction using LSTM+Optuna automatic optimization; automatically matching stable / volatile / small-sample specific models according to flower categories; dynamically weighting and fusing sub-model historical errors, combining regional and channel-specific weights to achieve price conversion across three ends, and constructing an error closed-loop iteration mechanism and engineering optimization. This invention achieves high-precision multi-period prediction with an accuracy ≥95%, adapting to 60+ mainstream flower categories, and supporting full-chain decision-making.

Description

Technical Field

[0001] This invention relates to the fields of big data time series mining, machine learning multi-model fusion, agricultural big data, and smart agriculture. Specifically, it relates to a method and system for predicting flower price index using multi-model adaptive fusion, which is applicable to price monitoring, supply chain decision-making, and planting planning in the flower industry. Background Technology

[0002] The flower industry is a high-value-added sector within my country's agricultural industry. Flower prices are influenced by multiple factors, including regional differences in production and sales areas, seasonal holidays, extreme weather, supply and demand, logistics costs, and variety specifications and grades. They exhibit strong seasonality, large cyclical fluctuations, significant sudden nonlinear fluctuations, and significant differences between varieties, placing extremely high demands on the accuracy and adaptability of price forecasting. Existing flower price forecasting technologies have the following significant shortcomings: (1) Model uniformity: Most of them adopt a single ARIMA or a single LSTM model, without distinguishing the differences in factor weights for short / medium / long-term predictions, and cannot adapt to the different patterns of short-term sudden shocks and long-term industry trends; lacking time series decomposition and multi-model fusion architecture, it is difficult to fit linear trends, seasonal cycles and nonlinear sudden fluctuations at the same time. (2) Lack of cycle adaptation: The difference in factor weights between short, medium and long-term forecasts is not distinguished. Short-term shocks such as extreme weather and sudden events cannot be amplified in short-term forecasts, and long-term forecasts cannot smooth out short-term noise and amplify long-term industry trends. The forecast accuracy fluctuates greatly with the cycle. (3) Insufficient category adaptation: Differentiated modeling was not performed for different flower categories. Stable green plants and fluctuating holiday gift flowers and small sample new varieties used the same model parameters, resulting in poor generalization ability and inability to adapt to the prediction needs of 60+ mainstream flower categories. (4) Data preprocessing generalization: Existing technologies for handling missing and outlier values ​​are general solutions that cannot adapt to the seasonal missing values ​​and peak price characteristics of the flower industry during holidays. They are prone to accidentally deleting reasonable fluctuations during holidays or failing to effectively supplement seasonal missing data such as the Spring Festival market closure. (5) Lack of self-iteration capability: The system lacks an error closed-loop feedback mechanism, the model parameters rely on manual debugging, and it cannot automatically backtest the prediction accuracy or automatically fine-tune the parameters. After long-term operation, the accuracy gradually decreases and the maintenance cost is extremely high. (6) The prediction results are black box outputs: the contribution of influencing factors and transmission links cannot be quantified, and there is no intelligent question answering and visualization decision-making ability. (7) Engineering performance bottleneck: The massive historical data uses full table batch query, which has problems such as query timeout, low CPU utilization and no task scheduling dependency, and cannot support the real-time prediction needs of large-scale full category.

[0003] Therefore, there is an urgent need for a flower price index prediction scheme that can adapt to the characteristics of the flower industry, support multi-model adaptive fusion, and have self-iterative capabilities, in order to solve the above-mentioned shortcomings of existing technologies. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies, such as single model, poor adaptation to cycle and category, generalized data preprocessing, lack of self-iteration capability, and insufficient engineering performance. It provides a multi-model adaptive fusion method and system for predicting flower price index, which integrates multi-source heterogeneous flower transaction data from offline flower markets, wholesale / retail platforms, and online supermarkets, and integrates influencing factors from all dimensions, including region, category, weather, holidays, logistics, production and sales, and public opinion, to achieve high-precision, adaptive, and self-optimizing prediction of flower price index.

[0005] This invention provides a multi-model adaptive fusion method for predicting flower price indices, comprising the following steps: S1. Differentiated Modeling and Weighting by Period: The forecasting period is divided into three categories: short-term (1-7 days), medium-term (8-14 days), and monthly / quarterly long-term. For short-term forecasting, the ARIMA model is used to amplify the weight of sudden factors such as extreme weather and short-term sales fluctuations, while weakening the weight of long-term industry trend factors. For medium-term forecasting, a composite model combining STL decomposition and random forest is used to fit seasonal cycles and nonlinear fluctuations. For long-term forecasting, the LSTM+Optuna automatic hyperparameter optimization model is used to automatically optimize the number of hidden layer neurons and the learning rate hyperparameter, smoothly reducing the weight of short-term factors such as extreme weather, and amplifying the weight of factors such as regional industrial development, long-term sales trends, and annual planting scale. S2. Adaptive Model Routing for Flower Categories: Based on the price fluctuation characteristics of flower categories, they are automatically divided into three categories: stable categories, volatile categories, and new varieties with small sample sizes. Different model parameters, adaptive decay factors, and price fluctuation alarm thresholds are then matched to each category. For stable categories, an improved ARIMA model is used, employing trend-weighted smoothing prediction and increasing the weight of recent price trends. For volatile categories, an LSTM or XGBoost model is used, employing a multi-factor nonlinear model and introducing adaptive decay factors, holiday influence factors, and price volatility differentiation parameters. For new varieties with small sample sizes, a GRNN generalized regression neural network is used for small-sample learning. S3. Multi-model dynamic weighted fusion and three-terminal price conversion: Based on the prediction results of each sub-model, the fusion weights are dynamically allocated according to the historical prediction MAPE error of each sub-model, and the fused wholesale price index prediction value of the place of origin is output; Based on the pre-maintained three-terminal price calculation model of flower category wholesale of place of origin, wholesale of place of destination, and retail of place of destination, combined with the parameters of circulation distance, cold chain logistics cost, and regional consumption preference, the weight of circulation cost factor is dynamically adjusted for different places of destination and different sales channels, and the wholesale price index of place of destination and the retail price index of place of destination are automatically derived from the wholesale price index of place of origin. S4. Error Closed Loop and Automatic Parameter Iteration: Periodically backtrack the predicted price index and the actual transaction price index to calculate the prediction accuracy of each category; when the accuracy is lower than the preset threshold, recalculate the flower feature values ​​and price ratio parameters, automatically analyze the causes of errors, fine-tune the model parameters and limit the adjustment range to within ±10%; if the accuracy deteriorates after adjustment, it automatically rolls back to the historical optimal parameters, forming a self-iterative closed loop of feature calculation - model prediction - accuracy backtesting - parameter optimization. S5. Intelligent Explanation and Price Anomaly Warning: Introduces SHAP value to quantify the price contribution of each factor, combines flower price knowledge graph to realize price impact chain explanation, and combines large language model to realize natural language question answering and explanation of the cause of flower price fluctuations; realizes daily price anomaly warning based on high / medium / low three-level thresholds for varieties, regions and circulation links. S6. Engineering-level data and scheduling optimization: Sliding window incremental data loading and batch small-granularity queries are used to replace the one-time full table big data query; multi-threaded concurrent execution of prediction tasks is adopted; a timed task orchestration is designed to realize the sequential scheduling of feature calculation, ratio analysis, threshold update, short-term prediction, medium-term prediction, and accuracy backtesting, and expired features are automatically marked and the prediction confidence is reduced. This invention also provides a multi-model adaptive fusion flower price index prediction system, comprising: an adaptive model routing module, a multi-model fusion prediction module, an error closed-loop iteration module, an intelligent interpretation and early warning module, and a parallel scheduling optimization module.

[0006] Compared with the prior art, the present invention has the following beneficial effects: (1) Adaptive model routing innovation: Automatically match the optimal model according to the prediction cycle and category characteristics to meet the prediction needs of 60+ mainstream flower categories and 8000+ flower subcategories. (2) Periodic differentiated weights and automatic hyperparameter optimization: Long-term prediction adopts LSTM+Optuna automatic hyperparameter optimization, which improves the parameter tuning efficiency by 80% and the long-term prediction accuracy by 2%. (3) Construction of industry-specific derivative features: Demand intensity index, supply risk index, etc. enhance nonlinear fitting ability and improve prediction accuracy by 3%. (4) Dynamic weighting by region and channel: cross-regional prediction accuracy improved by 4%. (5) Error feedback and automatic parameter fine-tuning mechanism: realizes unattended self-evolution of the model, greatly reducing maintenance costs. (6) Three-layer model dynamic fusion architecture: prediction accuracy is stable at ≥95%, and MAPE is controlled within 12%. (7) Enhanced interpretability of knowledge graph: Combine large models to realize natural language question answering and multi-level early warning. (8) Engineering performance optimization: Supports large-scale real-time prediction of all categories. (9) Full-chain price coverage: Simultaneously output price indices for wholesale at the place of origin, wholesale at the place of sale, and retail at the place of sale. Attached Figure Description

[0007] Figure 1 This is a diagram showing the overall module architecture of the system of the present invention;

[0008] Figure 2 This is a flowchart of the prediction process for multi-model adaptive fusion in this invention;

[0009] Figure 3 This is a logical diagram illustrating the category-adaptive model routing of the present invention;

[0010] Figure 4 This is a flowchart of the automatic iteration of error closed-loop in this invention. Detailed Implementation

[0011] The present invention will be further described in detail below with reference to specific embodiments.

[0012] Example 1: Mid-term forecasting of rose price index using multi-model adaptive fusion. This embodiment uses the mid-term forecast of the rose price index at Kunming Dounan Flower Market as an example to illustrate the implementation process of the present invention: 1. Data Preprocessing: Daily rose transaction data from Dounan Flower Market over the past two years were collected, and Kunming meteorological data, national holiday data, logistics cost data, monthly rose sales data, and social media popularity data for the corresponding periods were simultaneously aggregated; for missing values ​​due to the market closure during the Spring Festival, time-series interpolation was used to supplement the missing values ​​by combining the rose production estimates of the same period last year; 15-dimensional basic features including average price, volatility, sales volume, temperature, and holiday indicators were constructed, as well as exclusive derived features such as the pre-Valentine's Day demand intensity index (3.2 times that of weekdays) and the monthly supply risk index. 2. Time series decomposition: The STL algorithm is used to decompose the rose price time series into a long-term trend term (reflecting the long-term price trend of the rose industry), a seasonal cycle term (reflecting the cyclical fluctuations of peak and off-peak seasons), and an irregular residual term (reflecting the fluctuations of extreme weather and sudden demand). 3. Model matching: This forecast is a medium-term forecast for the next 10 days. Roses are a type of fluctuating holiday gift flower. The system automatically matches the STL + Random Forest composite model. Linear regression is used to fit the trend term, and random forest is used to fit the nonlinear effects of factors such as weather, holidays, and demand intensity for the residual term. 4. Fusion Prediction: Combining the historical prediction errors of the sub-models, the fusion weights are dynamically allocated, and the predicted value of the wholesale price index of the place of origin is output. Based on the category price ratio parameters of roses, combined with the 2,500-kilometer circulation distance from Kunming to Beijing and the cold chain logistics cost, the wholesale price index and retail price index of Beijing sales place are automatically calculated, and the weight of the logistics cost factor is automatically increased for Beijing sales place. 5. Results Verification: The prediction accuracy of this embodiment reaches 95%, and the MAPE is 8.7%, which is far better than the 90% accuracy of the traditional single ARIMA model.

[0013] Example 2: Automatic Parameter Iteration in Error Closed-Loop This embodiment illustrates the error closed-loop iteration mechanism of the present invention: 1. Accuracy backtesting: Every Monday, the system automatically backtests the predicted and actual transaction values ​​of the plant price index from the previous two weeks. The calculated prediction accuracy for the plant category is 87.2%, which is lower than the preset threshold of 90%. 2. Error Attribution: The system automatically identifies the error as originating from the underestimation of the impact weight of recent extreme high temperatures, which leads to the model failing to fully capture the price increase caused by the high temperatures in the transportation of green plants. 3. Parameter fine-tuning: The system automatically fine-tunes the weights of high-temperature meteorological factors by 8%, which does not exceed the ±10% limit. After retraining the model, the test accuracy is improved to 95%, and the system automatically saves the new parameters. 4. Rollback mechanism: If the test accuracy drops to 85% after adjustment, the system will automatically trigger the rollback mechanism to restore the historical optimal parameters before adjustment, thus avoiding model deterioration.

[0014] Example 3: Engineering-level scheduling optimization This embodiment illustrates the engineering optimization mechanism of the present invention: When performing monthly forecasting tasks for all product categories, the system uses a sliding window incremental query, reading only the valid data from the past two years, rather than the full 10 years of historical data, reducing the single query time from 12 seconds to 1.5 seconds. Simultaneously, it uses 8 threads to concurrently execute forecasting tasks (with a 12-core server CPU and threads configured to 2 / 3 of the CPU cores), reducing the forecasting time for all 60 product categories from 20 minutes to 4 minutes. Tasks are executed in dependency order, first completing feature calculations, then performing forecasts, and finally backtesting accuracy to ensure data consistency.

[0015] Example 4: Prediction of new varieties in small sample sizes This embodiment uses the price prediction of the popular new variety "pink pineapple" as an example to illustrate the small sample adaptability of the present invention: Pink pineapple is a new variety with only 800 transaction data points accumulated, representing a small sample scenario. The system automatically identifies this variety as a new variety with a small sample size and matches it with the GRNN generalized regression neural network model. By inputting the pineapple's category attributes, region, season, and meteorological characteristics, prediction can be completed without a large amount of historical data. The final prediction accuracy reached 86.3%, far superior to the 72% accuracy of the traditional LSTM model, solving the pain point of insufficient data for predicting new varieties.

[0016] Example 5: Extreme Weather Warning This embodiment illustrates the early warning mechanism of the present invention: When the system detects heavy rain warning data in the Kunming production area, it automatically activates the supply risk index and increases the weight of climate risk factors. If it predicts that the price of roses will rise by 18% in the next 3 days, exceeding the red warning threshold of 15%, the system automatically triggers a red warning and generates a link explanation through SHAP value and knowledge graph: "Extreme heavy rain → logistics interruption for 2 days → reduced supply → price increase of 18%". At the same time, it pushes the response suggestion "stock up in advance, and supplement local rose varieties to stabilize price fluctuations" to wholesalers and retailers 24 hours in advance.

Claims

1. A multi-model adaptive fusion method for predicting flower price indices, characterized in that, Includes the following steps: S1. Differentiated Modeling and Weighting by Period: The forecasting period is divided into three categories: short-term (1-7 days), medium-term (8-14 days), and monthly / quarterly long-term. For short-term forecasting, the ARIMA model is used to amplify the weight of sudden factors such as extreme weather and short-term sales fluctuations, while weakening the weight of long-term industry trend factors. For medium-term forecasting, a composite model combining STL decomposition and random forest is used to fit seasonal cycles and nonlinear fluctuations. For long-term forecasting, the LSTM+Optuna automatic hyperparameter optimization model is used to automatically optimize the number of hidden layer neurons and the learning rate hyperparameter, smoothly reducing the weight of short-term factors such as extreme weather, and amplifying the weight of factors such as regional industrial development, long-term sales trends, and annual planting scale. S2. Adaptive model routing for flower categories: Based on the price fluctuation characteristics of flower categories, they are automatically divided into three categories: stable categories, volatile categories, and new varieties with small sample sizes, and different model parameters, adaptive decay factors, and price fluctuation alarm thresholds are matched to each category respectively. For stable product categories, the ARIMA model is improved by using trend-weighted smoothing forecasting and increasing the weight of recent price trends. For volatile commodity categories, LSTM or XGBoost models are matched, and multi-factor nonlinear models are adopted, introducing adaptive decay factors, holiday impact factors, and price volatility differentiation parameters; for new commodities with small samples, GRNN generalized regression neural networks are matched for small sample learning; sliding windows, model hyperparameters, and price volatility alarm thresholds are automatically matched according to commodity categories. S3. Multi-model dynamic weighted fusion and three-terminal price conversion: Based on the prediction results of each sub-model, the fusion weights are dynamically allocated according to the historical prediction MAPE error of each sub-model, and the fused wholesale price index prediction value of the place of origin is output; Based on the pre-maintained three-terminal price calculation model of flower category wholesale of place of origin, wholesale of place of destination, and retail of place of destination, combined with circulation distance, cold chain logistics cost, and regional consumption preference parameters, the weight of circulation cost factor is dynamically adjusted for different places of destination and different sales channels, and the wholesale price index of place of destination and the retail price index of place of destination are automatically derived from the wholesale price index of place of origin; S4. Error Closed Loop and Automatic Parameter Iteration: Periodically backtrack the predicted price index and the actual transaction price index to calculate the prediction accuracy of each category; when the accuracy is lower than the preset threshold, recalculate the flower feature values ​​and price ratio parameters, automatically analyze the causes of error, fine-tune the model parameters and limit the adjustment range to within ±10%; if the accuracy deteriorates after adjustment, it will automatically roll back to the historical optimal parameters, forming a self-iterative closed loop of feature calculation - model prediction - accuracy backtesting - parameter optimization. S5. Intelligent Explanation and Price Anomaly Warning: Introduces SHAP value to quantify the price contribution of each factor, combines flower price knowledge graph to realize price impact chain explanation, and combines large language model to realize natural language question answering and explanation of the reasons for flower price fluctuations; realizes daily price anomaly warning based on high / medium / low three-level thresholds for varieties, regions and circulation links. S6. Engineering-level data and scheduling optimization: Sliding window incremental data loading and batch small-granularity queries are used to replace one-time full table big data queries; multi-threaded concurrent execution of prediction tasks is adopted, with the number of concurrent threads configured to be two-thirds to three-quarters of the number of server CPU cores; a scheduled task orchestration is designed to realize the sequential scheduling of feature calculation, ratio analysis, threshold update, short-term prediction, medium-term prediction, and accuracy backtesting, ensuring data freshness, automatically marking expired features and reducing prediction confidence.

2. A multi-model adaptive fusion flower price index prediction system, characterized in that, For performing the prediction method of claim 1, comprising: (1) Adaptive model routing module: It has built-in ARIMA, LSTM+Optuna, STL+random forest and GRNN model units, and automatically matches the corresponding model and parameter configuration according to the prediction period and flower category characteristics; (2) Multi-model fusion prediction module: used to dynamically allocate fusion weights based on the historical errors of sub-models, and combine regional and channel weights to complete the automatic conversion of the price index of the three ends of the production / sales wholesale and sales retail. (3) Error closed-loop iteration module: used for periodic backtesting of prediction accuracy, automatic error attribution, parameter fine-tuning and automatic rollback of deterioration, to realize model self-iterative optimization; (4) Intelligent Explanation and Early Warning Module: It is used to realize factor contribution and link explanation based on SHAP value and knowledge graph, and to realize natural language interaction by combining large model to complete multi-dimensional price fluctuation early warning; (5) Parallel scheduling optimization module: used to realize incremental batch data query, multi-threaded concurrent prediction, timed task dependency orchestration, and feature freshness verification.

3. The method according to claim 1, characterized in that, The short-term forecast takes daily time-series data from the past 1 to 3 months as input and outputs daily and weekly flower price indices; the long-term forecast takes long-term time-series data from the past 1 to 2 years as input and outputs monthly and quarterly flower price indices.

4. The method according to claim 1, characterized in that, The stable category includes foliage plants and potted green plants, while the volatile category includes holiday gift flowers such as roses, carnations, and lilies. Differentiated volatility and decay parameters are configured for the volatile category.

5. The method according to claim 1, characterized in that, The scheduled tasks are executed sequentially according to their dependencies: monthly feature value analysis, price ratio parameter update, risk threshold configuration, daily short-term forecast, daily medium-term forecast, and weekly accuracy backtesting and automatic parameter fine-tuning.

6. The system according to claim 2, characterized in that, The system supports plug-in model management and version rollback, and can add flower categories and coverage areas as needed.

7. The system according to claim 2, characterized in that, The system has a built-in price fluctuation early warning unit. When the predicted price fluctuation exceeds ±15%, a red alert is triggered, and response suggestions are automatically pushed.

8. The system according to claim 2, characterized in that, The system can simultaneously output price indices for wholesale at the place of origin, wholesale at the place of sale, and retail at the place of sale, adapting to the needs of multiple scenarios such as government supervision, planting planning, supply chain pricing, and retail replenishment.

9. The method according to claim 1, characterized in that, Before prediction, feature freshness is checked, expired features are marked with warnings and prediction confidence is reduced.

10. The method according to claim 1, characterized in that, For sales destinations in Beijing, the weight of the distribution distance factor is automatically increased; for sales destinations in Guangzhou, the weight of the climate risk factor is automatically increased; and for online channels, the weight of the logistics cost factor is automatically increased.

Images