Intelligent industry chain optimization method and system based on artificial intelligence

Abstract

The invention provides an intelligent industry chain optimization method and system based on artificial intelligence, and relates to the technical field of industry chain optimization. The method comprises the following steps: collecting historical industry chain data of an enterprise, constructing a training data set, and training to obtain a market demand prediction model; performing data segmentation on the historical industry chain data, inputting the data into the market demand prediction model, generating a plurality of market demand prediction values, and determining a first prediction fluctuation range; obtaining target industry chain data and inputting the target industry chain data into the market demand prediction model to generate a first demand prediction result; performing data fluctuation analysis and residual fluctuation analysis on the target industry chain data, and correcting the first predicted fluctuation range to generate a second abnormal fluctuation range; and generating a second demand prediction result based on the first demand prediction result and the second abnormal fluctuation range, and optimizing the industry chain of the enterprise according to the second demand prediction result. According to the invention, the intelligent level of enterprise industry chain optimization is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial chain optimization, and in particular to an intelligent industrial chain optimization method and system based on artificial intelligence. Background Art

[0002] In the current increasingly complex market environment, the accuracy and flexibility of industrial chain management are crucial to improving the competitiveness of enterprises. Industrial chain optimization involves multiple links such as market demand forecasting, inventory management, supply chain scheduling and production scheduling, and the links are highly linked and interdependent. Accurately predicting market demand and conducting intelligent industrial chain optimization can help companies effectively reduce costs, reduce resource waste and improve service levels. However, the management of the industrial chain depends on the analysis of multi-source data such as market demand, supply chain, and production progress. Some technologies do not consider the effective analysis and integration of multi-source data, as well as the complexity and variability of industrial chain data related data brought about by the increasingly complex market. The level of intelligence of enterprise industrial chain optimization needs to be improved. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes an intelligent industrial chain optimization method and system based on artificial intelligence, which adopts multi-source data fusion and combines with adaptive adjustment mechanism to realize the analysis and prediction of market demand, and incorporates demand fluctuations and uncertainties in complex market environments into the analysis and prediction process, thereby enhancing the intelligence level of industrial chain optimization, aiming to improve the efficiency and flexibility of enterprises in industrial chain management.

[0004] The first aspect of the present invention provides an intelligent industrial chain optimization method based on artificial intelligence, comprising:

[0005] Collect the company's historical industrial chain data, which includes market demand data, supply chain data, production data, and market environment data. Build a training data set based on the historical industrial chain data, and obtain a market demand forecasting model through training the training data set.

[0006] Perform data segmentation on historical industrial chain data to generate multiple groups of sub-sample data, and input the multiple groups of sub-sample data into the trained market demand forecasting model to generate market demand forecast values ​​for each group of sub-sample data;

[0007] Performing residual analysis on each group of sub-sample data based on the market demand forecast value, extracting residual features of each group of sub-sample data, and performing forecast fluctuation analysis based on multiple residual features to generate a first forecast fluctuation range;

[0008] Obtain the enterprise's current target industrial chain data and input it into the market demand forecasting model to generate the first demand forecasting result;

[0009] Performing data fluctuation analysis on the target industrial chain data to generate data fluctuation parameters, performing residual fluctuation analysis on the target industrial chain data to generate residual fluctuation parameters, and correcting the first predicted fluctuation range according to the data fluctuation parameters and the residual fluctuation parameters to generate a second abnormal fluctuation range;

[0010] generating a second demand forecast result based on the first demand forecast result and the second abnormal fluctuation range, and optimizing the enterprise's industrial chain according to the second demand forecast result;

[0011] Among them, the market demand forecasting model includes a multimodal data input layer, a local submodal processing layer, a multimodal data dynamic fusion layer, an LSTM layer and a fully connected output layer. The multimodal data input layer is used to receive the multimodal data in the training data set. The local submodal processing layer is used to extract the time series features of the multimodal data through multiple submodal processing networks. The multimodal data dynamic fusion layer is used to perform feature fusion on the time series features of the multimodal data to generate global fusion features. The LSTM layer is used to perform time series processing on the global fusion features to generate time series hidden features of multiple time steps. The fully connected output layer is used to generate market demand forecasting results based on multiple time series hidden features.

[0012] Preferably, the training of the market demand forecasting model includes:

[0013] After receiving the multimodal data in the training data set through the multimodal data input layer, the multimodal data is subjected to time series feature extraction through multiple submodal processing networks to generate local fusion features containing multiple modal time series features, including extracting market demand time series features, supply chain time series features and production time series features, and performing linear transformation processing on the market environment data through a feature fusion unit and then performing time step expansion to generate a non-time series feature sequence, and the non-time series feature sequence is subjected to feature splicing with the market demand time series features, supply chain time series features and production time series features to generate a local fusion feature;

[0014] The local fusion features are processed by the multimodal data dynamic fusion layer through the fully connected unit and the weighted fusion unit to generate the global fusion features, including determining the weight of each modal time series feature in the local fusion features through the fully connected unit, and fusing the multiple modal time series features based on the weight of each modal time series feature through the weighted fusion unit to generate the global fusion features, wherein:

[0015]

[0016] In the formula, is the global fusion feature, σ(·) is the Sigmoid function, is the local fusion feature, A and B are the weight matrix and bias term respectively, and They are market demand timing characteristics, supply chain timing characteristics, production timing characteristics, and non-timing characteristic sequences;

[0017] After generating the global fusion features, the long-term and short-term dependencies between multiple time steps are analyzed through the LSTM layer, and the output of the LSTM layer is processed through the fully connected output layer to output the predicted value of the market demand forecasting model. The mean square error loss function is used to analyze the difference between the output predicted value and the true value in the sample data set to generate the loss value of the market demand forecasting model. When the loss value reaches the preset loss range, the training of the market demand forecasting model is stopped to obtain a trained market demand forecasting model.

[0018] Preferably, the residual features of each group of sub-sample data are extracted, and a predicted fluctuation analysis is performed based on a plurality of residual features to generate a first predicted fluctuation range, including:

[0019] Based on the market demand forecast value, residual analysis is performed on each group of sub-sample data, the difference between the actual market demand value and the market demand forecast value of each group of sub-sample data is extracted, the demand residual characteristics of each group of sub-sample data are obtained, and a demand residual sequence containing the residual characteristics of multiple groups of sub-sample data is constructed;

[0020] The mean and standard deviation of the demand residual sequence are extracted. After the demand residual sequence is tested for normal distribution and confirmed to be in accordance with the normal distribution, the first forecast fluctuation range is determined based on the preset confidence level. The first forecast fluctuation range is μ t ±z d ·σ t , where μ t is the mean of the demand residual sequence, σ t is the standard deviation of the demand residual sequence, z d is the z value at the preset confidence level.

[0021] Preferably, performing data fluctuation analysis on the target industrial chain data to generate data fluctuation parameters, and performing residual fluctuation analysis on the target industrial chain data to generate residual fluctuation parameters include:

[0022] Determine the current time step t and the previous time step t-1 according to the target industry chain data, determine the first prediction value generated by the market demand prediction model at the current time step and the second prediction value generated by the market demand prediction model at the previous time step, and determine the data fluctuation parameter according to the first prediction value and the second prediction value;

[0023]

[0024] In the formula, w t is the data fluctuation parameter, y t ,y t-1are the first predicted value and the second predicted value respectively, ∈ is a constant term;

[0025] Determine the first prediction residual and the second prediction residual of the market demand forecasting model at time step t-1 and time step t-2 respectively, and determine the residual fluctuation parameter according to the first prediction residual and the second prediction residual:

[0026]

[0027] In the formula, r t is the residual fluctuation parameter, e t-1 、e t-2 are the first prediction residual and the second prediction residual respectively, and ε is a constant term.

[0028] Preferably, the second demand forecast result further includes:

[0029] Performing a fluctuation analysis on the target market demand forecast value in the first demand forecast result according to the second abnormal fluctuation range to obtain a market demand forecast range;

[0030] Among them, for the second abnormal fluctuation range, the first predicted fluctuation range is corrected and generated according to the data fluctuation parameter and the residual fluctuation parameter, and the second abnormal fluctuation range is μ t ±(w t +r t ) d ·σ t , for the market demand forecast range determined based on the second abnormal fluctuation range, the market demand forecast range is y t +μ t ±(w t +r t ) d ·σ t .

[0031] Preferably, one of the Shapiro-Wilk test and the Kolmogorov-Smirnov test is used to perform a normal distribution test on the demand residual sequence.

[0032] The second aspect of the present invention provides an intelligent industrial chain optimization system based on artificial intelligence, which is used to implement the above-mentioned intelligent industrial chain optimization method based on artificial intelligence, including:

[0033] The data collection module is used to collect the historical industrial chain data of the enterprise, which includes market demand data, supply chain data, production data and market environment data;

[0034] Model training module, used to build a training data set based on historical industrial chain data, and obtain a market demand forecasting model through training the training data set;

[0035] The sample prediction module is used to segment the historical industrial chain data to generate multiple groups of sub-sample data, and input the multiple groups of sub-sample data into the trained market demand prediction model to generate the market demand prediction value of each group of sub-sample data;

[0036] The market demand analysis module is used to obtain the current target industrial chain data of the enterprise and input it into the market demand forecasting model to generate the first demand forecasting result;

[0037] A forecast fluctuation analysis module is used to perform residual analysis on each group of sub-sample data based on the market demand forecast value, extract the residual features of each group of sub-sample data, and perform forecast fluctuation analysis based on multiple residual features to generate a first forecast fluctuation range;

[0038] A predicted fluctuation correction module is used to perform data fluctuation analysis on the target industrial chain data to generate data fluctuation parameters, perform residual fluctuation analysis on the target industrial chain data to generate residual fluctuation parameters, and correct the first predicted fluctuation range according to the data fluctuation parameters and the residual fluctuation parameters to generate a second abnormal fluctuation range;

[0039] The market demand correction module is used to generate a second demand forecast result based on the first demand forecast result and the second abnormal fluctuation range, and optimize the enterprise's industrial chain according to the second demand forecast result.

[0040] Preferably, the market demand forecasting model trained by the model training module further includes:

[0041] The market demand forecasting model includes a multimodal data input layer, a local submodal processing layer, a multimodal data dynamic fusion layer, an LSTM layer and a fully connected output layer. The multimodal data input layer is used to receive the multimodal data in the training data set. The local submodal processing layer is used to extract the time series features of the multimodal data through multiple submodal processing networks. The multimodal data dynamic fusion layer is used to perform feature fusion on the time series features of the multimodal data to generate global fusion features. The LSTM layer is used to perform time series processing on the global fusion features to generate time series hidden features of multiple time steps. The fully connected output layer is used to generate market demand forecasting results based on multiple time series hidden features.

[0042] The present invention has the following beneficial effects:

[0043] The present invention constructs a training data set through the collected historical industrial chain data, integrates multiple modal data such as market demand, supply chain, production and market environment, constructs and trains a market demand forecasting model that can process multimodal data, can consider the factors affecting market demand from multiple dimensions, capture the fluctuation trend of market demand, and can also incorporate the impact of supply chain fluctuations, production conditions, and external environment into the forecast, forming a more comprehensive forecasting model; residual analysis and fluctuation analysis are performed on the basis of the model prediction results, and on the basis of the first prediction fluctuation range generated based on historical data, a secondary correction is further performed in combination with the real-time fluctuation characteristics of the target data to ensure that the prediction interval corresponds closely to the actual demand changes, can automatically adapt to the short-term fluctuations and extreme situations of the market, and adjust the prediction interval to a more reasonable range, ensuring that the demand forecast results are closer to the actual changes in the market. Ultimately, through intelligent demand forecasting and uncertainty control, the inventory, production and supply chain management of the enterprise can be accurately optimized, and the overall responsiveness and resource utilization efficiency of the industrial chain are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A flowchart of an intelligent industrial chain optimization method based on artificial intelligence provided in an embodiment of the present invention.

[0045] Figure 2 A schematic diagram of the structure of an intelligent industrial chain optimization system based on artificial intelligence provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make those skilled in the art better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0047] like Figure 1 As shown, the embodiment of the present invention provides an intelligent industrial chain optimization method based on artificial intelligence, which specifically includes the following steps:

[0048] Step S10: Collect the historical industrial chain data of the enterprise, build a training data set based on the historical industrial chain data, and obtain a market demand forecasting model through training the training data set.

[0049] Specifically, first collect the company's historical industrial chain data, including but not limited to market demand data, supply chain data, production data and market environment data, etc. These data cover the key links of the company's operations and reflect the volatility of market demand, the stability of the supply chain, the allocation of production resources and the impact of the external market environment. After preprocessing the collected historical data, including data cleaning, outlier processing, normalization and other steps, the preprocessed historical industrial chain data is divided into training data sets to ensure that the data features of various dimensions such as market demand, supply chain, production and market environment can fully participate in the training. Then input the training data set into the market demand forecasting model to train the market demand forecasting model, so as to obtain a model that can accurately predict market demand.

[0050] Step S20: segment the historical industrial chain data to generate multiple groups of sub-sample data, input the multiple groups of sub-sample data into the trained market demand forecasting model respectively, and generate market demand forecast values ​​for each group of sub-sample data.

[0051] Specifically, the historical industrial chain data is further divided into multiple groups of sub-sample data in order to forecast and analyze market demand at a finer granularity. Each group of sub-sample data can cover different time periods, different market environment characteristics, etc., thus forming a variety of forecasting scenarios. Multiple groups of sub-sample data are input into the trained market demand forecasting model to obtain the market demand forecast value of each group of sub-sample data. These forecast values ​​reflect the changing trend of market demand under different sample scenarios.

[0052] Step S30: perform residual analysis on each group of sub-sample data based on the market demand forecast value, extract the residual characteristics of each group of sub-sample data, perform forecast fluctuation analysis based on multiple residual characteristics to generate a first forecast fluctuation range, obtain the company's current target industrial chain data and input it into the market demand forecast model to generate a first demand forecast result.

[0053] Specifically, residual analysis is performed on each group of sub-sample data based on multiple market demand forecast values ​​generated by the model, that is, the actual market demand of each group of sub-sample data is compared with the forecast value, and the residual of each sub-sample is calculated to help identify the error characteristics of the model prediction based on the prediction deviation of the model in different scenarios. Then, feature extraction is performed on the residual of each group of sub-sample data, such as the mean, variance, and rate of change of the residual, so as to evaluate the fluctuation characteristics of the prediction error and facilitate the identification of the error law and uncertainty source of the model under different samples. Based on multiple residual features, the prediction fluctuation analysis is performed, and the fluctuation range and change trend of the residual are determined by statistical analysis, thereby generating a first prediction fluctuation range. This fluctuation range provides a preliminary uncertainty interval for the prediction result, indicating the fluctuation range of market demand under different sub-sample scenarios. After obtaining the current target industrial chain data of the enterprise, that is, the current market demand, supply chain, production and market environment data, it is input into the trained market demand forecasting model to generate the first demand forecast result of the current market demand.

[0054] Step S40: Perform data fluctuation analysis on the target industrial chain data to generate data fluctuation parameters, perform residual fluctuation analysis on the target industrial chain data to generate residual fluctuation parameters, and modify the first predicted fluctuation range according to the data fluctuation parameters and the residual fluctuation parameters to generate a second abnormal fluctuation range.

[0055] Specifically, data fluctuation analysis is performed on the target industrial chain data to generate data fluctuation parameters that comprehensively characterize the short-term change characteristics of market demand data, the stability of supply chain data, the adjustment frequency of production data and other fluctuation characteristics to reflect the fluctuation characteristics of the current industrial chain status. Based on the change characteristics of historical residuals, residual fluctuation analysis is performed on the current target industrial chain data to generate residual fluctuation parameters. Residual fluctuation parameters are used to describe the error level and volatility of the current forecast and help judge the reliability of the forecast results. Then, the first forecast fluctuation range is corrected by combining the data fluctuation parameters and residual fluctuation parameters to generate the second abnormal fluctuation range. This range dynamically adjusts the fluctuation range of the initial forecast to make the forecast range closer to the current data characteristics, thereby improving the robustness and adaptability of the forecast. The abnormal fluctuation range can better reflect the current market conditions and potential changes in forecast demand.

[0056] Step S50: Generate a second demand forecast result based on the first demand forecast result and the second abnormal fluctuation range, and optimize the enterprise's industrial chain according to the second demand forecast result.

[0057] Specifically, based on the first demand forecast result and the second abnormal fluctuation range, a more accurate second demand forecast result is generated. It is worth noting that the conventional time series forecasting model assumes that the input data is deterministic, that is, it assumes that historical data represents future trends. The first demand forecast result under this method is a deterministic forecast result generated based on the model, which may not be accurate enough in a real environment because data in the real world often has a high degree of uncertainty. For example, supplier delivery delays, sudden fluctuations in market demand, sudden equipment failures, etc., will have a significant impact on the stability of the industrial chain. Taking this factor into account, by analyzing the residual characteristics of the historical industrial chain data, a first predicted fluctuation range is generated to characterize the fluctuation impact of the fluctuation characteristics of the historical industrial chain data on the model's prediction. Further data fluctuation analysis and data residual analysis are performed on the collected target industrial chain data, that is, considering the deviation between the currently collected data and the historical data, the first predicted fluctuation range that reflects the law of historical data is corrected to obtain the second predicted fluctuation range, and then the first demand forecast result generated by the model is analyzed according to the second predicted fluctuation range to obtain a second demand forecast result that is a more reliable forecast of market demand, which includes the fluctuation of the predicted value. Enterprises can optimize the industrial chain based on the second demand forecast result, such as adjusting inventory levels according to demand forecast results to ensure sufficient supply within the uncertainty fluctuation range, avoid insufficient inventory or excessive backlogs, flexibly adjust production volume according to predicted demand, reduce resource waste, and improve production efficiency. According to the forecast results, the material, logistics and procurement plans of the supply chain are optimized to ensure the stability and timeliness of the supply chain. Response strategies are formulated in advance for predicted market demand fluctuations, such as market promotions, pricing adjustments, etc., to ensure that enterprises can respond in a timely manner when demand changes and improve their market competitiveness.

[0058] The intelligent industrial chain optimization method provided by the present invention enables market demand forecasting to have higher accuracy and real-time performance through full-process control from data collection, model training, residual analysis to uncertainty interval correction. On the basis of the first predicted fluctuation range generated based on historical data, a secondary correction is further performed in combination with the real-time fluctuation characteristics of the target data to ensure that the predicted interval closely corresponds to the actual demand changes. Ultimately, through intelligent demand forecasting and uncertainty control, the company's inventory, production and supply chain management can be accurately optimized, thereby improving the overall responsiveness and resource utilization efficiency of the industrial chain.

[0059] In one implementation process, for the market demand prediction model in step S10, its structure includes a multimodal data input layer, a local sub-modal processing layer, a multimodal data dynamic fusion layer, an LSTM layer and a fully connected output layer.

[0060] The multimodal data input layer is used to receive multimodal data in the training data set, including market demand data, supply chain data, production data, and market environment data. Through the input of this layer, it can ensure that multi-dimensional data features enter the model, providing rich contextual information for subsequent analysis.

[0061] The local submodal processing layer is used to extract the time series features of multimodal data through multiple submodal processing networks. These submodal networks perform specialized processing on different types of data to generate data features suitable for each modality. The advantage of the local submodal processing layer is that it can independently analyze the data features of each modality and extract specific features of market demand, supply chain fluctuations, and production conditions, including market demand time series features, supply chain time series features, and production time series features. These features will then enter the next layer for fusion.

[0062] The multimodal data dynamic fusion layer is used to perform feature fusion on the time series features of multimodal data to generate global fusion features. The LSTM layer is used to perform time series processing on the global fusion features, with the main purpose of capturing long-term and short-term dependencies. This layer processes input features of multiple time steps and learns the hidden relationships between different time steps. Through long short-term memory units, the model can analyze historical demand trends to better predict future market demand. After generating time series hidden features of multiple time steps, the output of the LSTM layer is also passed to the fully connected layer for further processing.

[0063] The fully connected output layer is used to generate market demand forecast results based on multiple time series hidden features. It is responsible for nonlinear transformation of the features output by the LSTM layer to generate the final forecast value of the market demand forecast model. This forecast value is the future market demand predicted based on the input data.

[0064] In the process of training the market demand forecasting model through the training data set, after the multimodal data input layer receives the multimodal data in the training data set, the local submodal processing layer extracts time series features of the multimodal data through multiple submodal processing networks. Each submodal processing network is specifically a separate sub-LSTM unit, which is used to extract the dynamic features of the time series of multiple modal data and generate a hidden state of a time step.

[0065] For market demand time series data, supply chain time series data and production time series data, such as sales volume, inventory level, production capacity and other data, they are processed through the sub-modal processing network, and finally a local fusion feature containing multiple modal time series features is generated, including the extraction of market demand time series features, supply chain time series features and production time series features. For external environment non-time series data, such as weather, holidays and other data, the market environment data is linearly transformed by the feature fusion unit and then expanded in time steps to generate a non-time series feature sequence. In this embodiment, the numerical non-time series data is converted into a vector of a fixed dimension through a fully connected layer, that is, it is converted into a vector suitable for fusion with time series data through a linear transformation. Since the non-time series data has no time dependence, it needs to be extended to each time step so that it can be input into the model together with the time series data. Specifically, the non-time series features are copied to each time step and combined with the time series data of each time step to obtain a non-time series feature sequence. Finally, the non-time series feature sequence is feature spliced ​​with the market demand time series features, the supply chain time series features and the production time series features to generate a local fusion feature.

[0066] After generating local fusion features, the local fusion features are processed by the multimodal data dynamic fusion layer to generate global fusion features. The multimodal data dynamic fusion layer implements this process through fully connected units and weighted fusion units. Specifically, the weight of each modal time series feature in the local fusion feature is determined by the fully connected unit, that is, the local fusion feature is input into a fully connected layer, and each modal time series feature is converted into a weight to obtain the weight of each modal time series feature. Then, the weighted fusion unit is used to fuse multiple modal time series features based on the weight of each modal time series feature to generate a global fusion feature. In this process:

[0067]

[0068] In the formula, is the global fusion feature, σ(·) is the Sigmoid function, is the local fusion feature, A and B are the weight matrix and bias term respectively, and They are market demand timing characteristics, supply chain timing characteristics, production timing characteristics and non-timing characteristic sequences.

[0069] The global fusion features generated by the multimodal data dynamic fusion layer are input into the core LSTM layer. The LSTM layer captures the long-term and short-term dependencies between multiple time steps through memory units, obtains the time series hidden features of each time step, and generates the final time series output. Finally, it is input into the fully connected output layer, and the output of the LSTM layer is processed by the fully connected output layer, including nonlinear transformation of the features output by the LSTM layer, to generate the final prediction value of the market demand forecasting model, which is the future market demand predicted based on the input data.

[0070] In the process of model training, the mean square error loss function is used to analyze the difference between the output prediction value and the true value in the sample data set to generate the loss value of the market demand forecasting model. When the loss value reaches the preset loss range, the training of the market demand forecasting model is stopped to obtain a stable market demand forecasting model. The above market demand forecasting model uses the dynamic fusion of multimodal data and the time series feature analysis of LSTM to achieve more accurate market demand forecasting and ensure the stability and reliability of the forecasting results in a changing market environment.

[0071] In one implementation process, for step S30, the residual features of each group of sub-sample data are extracted, and a forecast fluctuation analysis is performed based on multiple residual features to generate a first forecast fluctuation range, specifically including:

[0072] Based on the market demand forecast value, residual analysis is performed on each group of sub-sample data to extract the difference between the actual market demand value and the market demand forecast value of each group of sub-sample data, and the demand residual characteristics of each group of sub-sample data are obtained, which represent the prediction error of the model on different sub-sample data. This can help identify the distribution characteristics of prediction deviations and errors. By aggregating the residual characteristics of multiple groups of sub-sample data, a demand residual sequence containing the residual characteristics of multiple groups of sub-sample data is constructed to reflect the random fluctuation characteristics that may exist in market demand forecasts.

[0073] After generating the demand residual sequence, the key features of the demand residual sequence are extracted. In this embodiment, the mean and standard deviation of the demand residual sequence are extracted. In order to ensure the rationality of the predicted fluctuation range, the demand residual sequence is first subjected to a normal distribution test to verify whether the residual sequence conforms to the normal distribution. The Shapiro-Wilk test, Kolmogorov-Smirnov test and other methods can be used to evaluate the distribution characteristics of the residual sequence, such as the Shapiro-Wilk test for normal distribution test. It is not specifically limited here. After the demand residual sequence is subjected to a normal distribution test and it is determined that the demand residual sequence conforms to the normal distribution, the first predicted fluctuation range is determined based on the preset confidence level. The first predicted fluctuation range is μ t ±z d ·σt , where μ t is the mean of the demand residual sequence, σ t is the standard deviation of the demand residual sequence, z d is the z value under the preset confidence level. For example, when the preset confidence level is 95%, z d The above method provides a preliminary credible interval for subsequent forecasts, indicating the possible upper and lower limits of market demand forecasts, which facilitates enterprises to control and optimize risks in resource allocation and production decisions.

[0074] In one implementation process, for step S40, the process of performing data fluctuation analysis on the target industrial chain data to generate data fluctuation parameters and performing residual fluctuation analysis on the target industrial chain data to generate residual fluctuation parameters specifically includes:

[0075] The data volatility parameter is mainly used to quantify the short-term volatility of the target industry chain data, so as to help adjust the dynamics of the forecast interval. Considering that the target industry chain data directly collected is relatively complex, it is difficult to directly analyze its fluctuations compared with historical data. The volatility of the predicted value over time can often indirectly reflect the uncertainty of the data. Therefore, the volatility of the predicted value is directly used to characterize the volatility of the data. The calculation is based on the rate of change of the market demand forecast value between the current time step and the previous time step. If the predicted value changes greatly in a short period of time, it means that the model may have a high degree of uncertainty.

[0076] Specifically, the current time step t and the previous time step t-1 are determined according to the target industrial chain data, the first prediction value generated by the market demand prediction model at the current time step and the second prediction value generated by the market demand prediction model at the previous time step are determined, and the data fluctuation parameter is determined according to the first prediction value and the second prediction value;

[0077]

[0078] In the formula, w t is the data fluctuation parameter, y t ,y t-1 are the first predicted value and the second predicted value, respectively, and ∈ is a constant term. The larger the value of the data fluctuation parameter, the more obvious the change of the current predicted value compared with the previous time step, indicating that the uncertainty of market demand increases, which indirectly reflects that the uncertainty of the collected data is large.

[0079] The residual fluctuation parameter is used to evaluate the volatility of the model's prediction error at different time steps. Considering that the true value corresponding to the predicted value at the current time step is still unknown, the changing characteristics of the historical residuals are used to characterize the changing characteristics of the current data. The prediction residual of the model is extracted from the historical data, that is, the difference between the predicted value of the model in the past time step and the actual true value. Finally, the residual fluctuation parameter is calculated to correct the uncertainty interval of the prediction.

[0080] Specifically, the first prediction residual and the second prediction residual of the market demand forecasting model at time step t-1 and time step t-2 are determined respectively, and the residual fluctuation parameter is determined according to the first prediction residual and the second prediction residual:

[0081]

[0082] In the formula, r t is the residual fluctuation parameter, e t-1 、e t-2 are the first prediction residual and the second prediction residual, respectively, and ε is a constant term. A larger value of the residual fluctuation parameter indicates a higher fluctuation in the prediction error, indicating that the model prediction is less stable at these time steps.

[0083] In order to more accurately reflect the uncertainty of current market demand, the present invention introduces a second abnormal fluctuation range, and based on this range, the preliminary market demand forecast value is corrected to generate a more reliable demand forecast interval. After determining the data fluctuation parameter and the residual fluctuation parameter in the above manner, the first abnormal fluctuation range is corrected by the data fluctuation parameter and the residual fluctuation parameter to generate the second abnormal fluctuation range.

[0084] The second abnormal fluctuation range is generated by correcting the first predicted fluctuation range according to the data fluctuation parameter and the residual fluctuation parameter. The second abnormal fluctuation range is μ t ±(w t +r t ) d ·σ t The data fluctuation parameter is used to adjust the dynamics of the market demand forecast value, and the residual fluctuation parameter is used to evaluate the fluctuation of the model prediction error. Considering the fluctuation of data fluctuation and prediction error at the same time, the upper and lower limits of the prediction interval are more flexible and can cover the extreme situations that may occur in market demand.

[0085] After determining the second abnormal fluctuation range, the market demand forecast range is determined based on the second abnormal fluctuation range, where the market demand forecast range is y t +μ t ±(w t +r t ) d ·σ t .yt The first forecast value generated by the market demand forecast model at the current time step, that is, the market demand forecast value generated by the model in the first demand forecast result. By setting the forecast range of the market demand forecast value within a dynamically adjusted interval in the above manner, it is ensured that the demand forecast not only takes into account market fluctuations, but also integrates the volatility of the forecast error, so that the forecast results can cover more market changes. This dynamic forecast interval setting helps enterprises to reserve sufficient safety space in actual decision-making and improve the risk resistance of industrial chain management.

[0086] like Figure 2 As shown, the embodiment of the present invention further provides an intelligent industrial chain optimization system based on artificial intelligence, which is used to implement the above-mentioned intelligent industrial chain optimization method based on artificial intelligence, including:

[0087] The data collection module is used to collect the historical industrial chain data of the enterprise, which includes market demand data, supply chain data, production data and market environment data;

[0088] Model training module, used to build a training data set based on historical industrial chain data, and obtain a market demand forecasting model through training the training data set;

[0089] Among them, the market demand forecasting model includes a multimodal data input layer, a local submodal processing layer, a multimodal data dynamic fusion layer, an LSTM layer and a fully connected output layer. The multimodal data input layer is used to receive the multimodal data in the training data set. The local submodal processing layer is used to extract the time series features of the multimodal data through multiple submodal processing networks. The multimodal data dynamic fusion layer is used to perform feature fusion on the time series features of the multimodal data to generate global fusion features. The LSTM layer is used to perform time series processing on the global fusion features to generate time series hidden features of multiple time steps. The fully connected output layer is used to generate market demand forecasting results based on multiple time series hidden features.

[0090] The training of the market demand forecasting model includes:

[0091] After receiving the multimodal data in the training data set through the multimodal data input layer, the multimodal data is subjected to time series feature extraction through multiple submodal processing networks to generate local fusion features containing multiple modal time series features, including extracting market demand time series features, supply chain time series features and production time series features, and performing linear transformation processing on the market environment data through a feature fusion unit and then performing time step expansion to generate a non-time series feature sequence, and the non-time series feature sequence is subjected to feature splicing with the market demand time series features, supply chain time series features and production time series features to generate a local fusion feature;

[0092] The local fusion features are processed by the multimodal data dynamic fusion layer through the fully connected unit and the weighted fusion unit to generate the global fusion features, including determining the weight of each modal time series feature in the local fusion features through the fully connected unit, and fusing the multiple modal time series features based on the weight of each modal time series feature through the weighted fusion unit to generate the global fusion features, wherein:

[0093]

[0094] In the formula, is the global fusion feature, σ(·) is the Sigmoid function, is the local fusion feature, A and B are the weight matrix and bias term respectively, and They are market demand timing characteristics, supply chain timing characteristics, production timing characteristics, and non-timing characteristic sequences;

[0095] After generating the global fusion features, the long-term and short-term dependencies between multiple time steps are analyzed through the LSTM layer, and the output of the LSTM layer is processed through the fully connected output layer to output the predicted value of the market demand forecasting model. The mean square error loss function is used to analyze the difference between the output predicted value and the true value in the sample data set to generate the loss value of the market demand forecasting model. When the loss value reaches the preset loss range, the training of the market demand forecasting model is stopped to obtain a trained market demand forecasting model.

[0096] The sample prediction module is used to segment the historical industrial chain data to generate multiple groups of sub-sample data, and input the multiple groups of sub-sample data into the trained market demand prediction model to generate the market demand prediction value of each group of sub-sample data;

[0097] The market demand analysis module is used to obtain the current target industrial chain data of the enterprise and input it into the market demand forecasting model to generate the first demand forecasting result;

[0098] A forecast fluctuation analysis module is used to perform residual analysis on each group of sub-sample data based on the market demand forecast value, extract the residual features of each group of sub-sample data, and perform forecast fluctuation analysis based on multiple residual features to generate a first forecast fluctuation range;

[0099] A predicted fluctuation correction module is used to perform data fluctuation analysis on the target industrial chain data to generate data fluctuation parameters, perform residual fluctuation analysis on the target industrial chain data to generate residual fluctuation parameters, and correct the first predicted fluctuation range according to the data fluctuation parameters and the residual fluctuation parameters to generate a second abnormal fluctuation range;

[0100] The market demand correction module is used to generate a second demand forecast result based on the first demand forecast result and the second abnormal fluctuation range, and optimize the enterprise's industrial chain according to the second demand forecast result.

[0101] The above is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. An intelligent industrial chain optimization method based on artificial intelligence, characterized in that: include: Collect the company's historical industrial chain data, which includes market demand data, supply chain data, production data, and market environment data. Build a training data set based on the historical industrial chain data, and obtain a market demand forecasting model through training the training data set. Perform data segmentation on historical industrial chain data to generate multiple groups of sub-sample data, and input the multiple groups of sub-sample data into the trained market demand forecasting model to generate market demand forecast values ​​for each group of sub-sample data; Performing residual analysis on each group of sub-sample data based on the market demand forecast value, extracting residual features of each group of sub-sample data, and performing forecast fluctuation analysis based on multiple residual features to generate a first forecast fluctuation range; Obtain the enterprise's current target industrial chain data and input it into the market demand forecasting model to generate the first demand forecasting result; Performing data fluctuation analysis on the target industrial chain data to generate data fluctuation parameters, performing residual fluctuation analysis on the target industrial chain data to generate residual fluctuation parameters, and correcting the first predicted fluctuation range according to the data fluctuation parameters and the residual fluctuation parameters to generate a second abnormal fluctuation range; generating a second demand forecast result based on the first demand forecast result and the second abnormal fluctuation range, and optimizing the enterprise's industrial chain according to the second demand forecast result; Among them, the market demand forecasting model includes a multimodal data input layer, a local submodal processing layer, a multimodal data dynamic fusion layer, an LSTM layer and a fully connected output layer. The multimodal data input layer is used to receive the multimodal data in the training data set. The local submodal processing layer is used to extract the time series features of the multimodal data through multiple submodal processing networks. The multimodal data dynamic fusion layer is used to perform feature fusion on the time series features of the multimodal data to generate global fusion features. The LSTM layer is used to perform time series processing on the global fusion features to generate time series hidden features of multiple time steps. The fully connected output layer is used to generate market demand forecasting results based on multiple time series hidden features.

2. According to claim 1, the intelligent industrial chain optimization method based on artificial intelligence is characterized in that: The training of the market demand forecasting model includes: After receiving the multimodal data in the training data set through the multimodal data input layer, the multimodal data is subjected to time series feature extraction through multiple submodal processing networks to generate local fusion features containing multiple modal time series features, including extracting market demand time series features, supply chain time series features and production time series features, and performing linear transformation processing on the market environment data through a feature fusion unit and then performing time step expansion to generate a non-time series feature sequence, and the non-time series feature sequence is subjected to feature splicing with the market demand time series features, supply chain time series features and production time series features to generate a local fusion feature; The local fusion features are processed by the multimodal data dynamic fusion layer through the fully connected unit and the weighted fusion unit to generate the global fusion features, including determining the weight of each modal time series feature in the local fusion features through the fully connected unit, and fusing the multiple modal time series features based on the weight of each modal time series feature through the weighted fusion unit to generate the global fusion features, wherein: In the formula, is the global fusion feature, σ(·) is the Sigmoid function, is the local fusion feature, A and B are the weight matrix and bias term respectively, and They are market demand timing characteristics, supply chain timing characteristics, production timing characteristics, and non-timing characteristic sequences; After generating the global fusion features, the long-term and short-term dependencies between multiple time steps are analyzed through the LSTM layer, and the output of the LSTM layer is processed through the fully connected output layer to output the predicted value of the market demand forecasting model. The mean square error loss function is used to analyze the difference between the output predicted value and the true value in the sample data set to generate the loss value of the market demand forecasting model. When the loss value reaches the preset loss range, the training of the market demand forecasting model is stopped to obtain a trained market demand forecasting model.

3. According to claim 1, the intelligent industrial chain optimization method based on artificial intelligence is characterized in that: The residual features of each group of sub-sample data are extracted, and a forecast fluctuation analysis is performed based on multiple residual features to generate a first forecast fluctuation range, including: Based on the market demand forecast value, residual analysis is performed on each group of sub-sample data, the difference between the actual market demand value and the market demand forecast value of each group of sub-sample data is extracted, the demand residual characteristics of each group of sub-sample data are obtained, and a demand residual sequence containing the residual characteristics of multiple groups of sub-sample data is constructed; The mean and standard deviation of the demand residual sequence are extracted. After the demand residual sequence is tested for normal distribution and confirmed to be in accordance with the normal distribution, the first forecast fluctuation range is determined based on the preset confidence level. The first forecast fluctuation range is μ t ±z d ·σ t , where μ t is the mean of the demand residual sequence, σ t is the standard deviation of the demand residual series, z d is the z value at the preset confidence level.

4. The method for optimizing an intelligent industrial chain based on artificial intelligence according to claim 1, characterized in that: Perform data fluctuation analysis on the target industrial chain data to generate data fluctuation parameters, and perform residual fluctuation analysis on the target industrial chain data to generate residual fluctuation parameters, including: Determine the current time step t and the previous time step t-1 according to the target industry chain data, determine the first prediction value generated by the market demand prediction model at the current time step and the second prediction value generated by the market demand prediction model at the previous time step, and determine the data fluctuation parameter according to the first prediction value and the second prediction value; In the formula, w t is the data fluctuation parameter, y t ,y t-1 are the first predicted value and the second predicted value respectively, ∈ is a constant term; Determine the first prediction residual and the second prediction residual of the market demand forecasting model at time step t-1 and time step t-2 respectively, and determine the residual fluctuation parameter according to the first prediction residual and the second prediction residual: In the formula, r t is the residual fluctuation parameter, e t-1 、e t-2 are the first prediction residual and the second prediction residual respectively, and ε is a constant term.

5. The method for optimizing an intelligent industrial chain based on artificial intelligence according to claim 4, characterized in that: For the second demand forecast result, it also includes: Performing a fluctuation analysis on the target market demand forecast value in the first demand forecast result according to the second abnormal fluctuation range to obtain a market demand forecast range; Among them, for the second abnormal fluctuation range, the first predicted fluctuation range is corrected and generated according to the data fluctuation parameter and the residual fluctuation parameter, and the second abnormal fluctuation range is μ t ±(w t +r t ) d ·σ t , for the market demand forecast range determined based on the second abnormal fluctuation range, the market demand forecast range is y t +μ t ±(w t +r t ) d ·σ t .

6. The method for optimizing an intelligent industrial chain based on artificial intelligence according to claim 5, characterized in that: The demand residual sequence is tested for normal distribution using either the Shapiro-Wilk test or the Kolmogorov-Smirnov test.

7. An intelligent industrial chain optimization system based on artificial intelligence, characterized in that: The system is used to implement an intelligent industrial chain optimization method based on artificial intelligence as described in any one of claims 1 to 6, including: The data collection module is used to collect the historical industrial chain data of the enterprise, which includes market demand data, supply chain data, production data and market environment data; Model training module, used to build a training data set based on historical industrial chain data, and obtain a market demand forecasting model through training the training data set; The sample prediction module is used to segment the historical industrial chain data to generate multiple groups of sub-sample data, and input the multiple groups of sub-sample data into the trained market demand prediction model to generate the market demand prediction value of each group of sub-sample data; The market demand analysis module is used to obtain the current target industrial chain data of the enterprise and input it into the market demand forecasting model to generate the first demand forecasting result; A forecast fluctuation analysis module is used to perform residual analysis on each group of sub-sample data based on the market demand forecast value, extract the residual features of each group of sub-sample data, and perform forecast fluctuation analysis based on multiple residual features to generate a first forecast fluctuation range; A predicted fluctuation correction module is used to perform data fluctuation analysis on the target industrial chain data to generate data fluctuation parameters, perform residual fluctuation analysis on the target industrial chain data to generate residual fluctuation parameters, and correct the first predicted fluctuation range according to the data fluctuation parameters and the residual fluctuation parameters to generate a second abnormal fluctuation range; The market demand correction module is used to generate a second demand forecast result based on the first demand forecast result and the second abnormal fluctuation range, and optimize the enterprise's industrial chain according to the second demand forecast result.

8. The intelligent industrial chain optimization system based on artificial intelligence according to claim 7 is characterized in that: For the market demand forecasting model trained by the model training module, it also includes: The market demand forecasting model includes a multimodal data input layer, a local submodal processing layer, a multimodal data dynamic fusion layer, an LSTM layer and a fully connected output layer. The multimodal data input layer is used to receive the multimodal data in the training data set. The local submodal processing layer is used to extract the time series features of the multimodal data through multiple submodal processing networks. The multimodal data dynamic fusion layer is used to perform feature fusion on the time series features of the multimodal data to generate global fusion features. The LSTM layer is used to perform time series processing on the global fusion features to generate time series hidden features of multiple time steps. The fully connected output layer is used to generate market demand forecasting results based on multiple time series hidden features.

Images