Ensemble Neural Network Forecasting for Spend Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing forecasting solutions for spend data management face inaccuracies, especially during abrupt deviations like the coronavirus pandemic, due to their inability to account for nonlinear pricing and non-continuous time series data, leading to high error rates and manual intervention in procurement and inventory management.
Innovation Solution
The implementation of a customized forecasting solution using dynamically adaptive weights for ensemble neural network architecture, which incorporates multiple multi-step ensemble models, data augmentation, and normalization to create a processed dataset that accurately predicts price, demand, and supply, accounting for volatile aspects and sudden changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If traditional statistical forecasting methods are used, then the forecasting process is simple and easy to implement, but the prediction accuracy deteriorates when there is abrupt deviation in data due to drastically changed environment
Solution Approach 1:
The patent transforms the forecasting approach by changing from traditional statistical parameters to neural network parameters that can adapt to non-linear relationships and abrupt environmental changes. The system uses multiple neural networks with different architectures (LSTM, GRU, Dense) and dynamically adjusts their weights based on performance, enabling accurate predictions during pandemic conditions while maintaining implementation feasibility through automated model selection.
Solution Approach 2:
The patent creates a composite forecasting system by combining multiple neural network architectures (LSTM, GRU, Dense networks) into an ensemble model. Each network type contributes different strengths to handle various patterns in the data, and their predictions are weighted and combined to achieve superior accuracy compared to any single model, particularly during abrupt deviations like the coronavirus pandemic.
2Measurement precision
If neural network architecture is used to handle nonlinear pricing and demand, then the model can capture complex patterns, but it leads to over-fitted models for materials with fluctuations in price and demand
Solution Approach 1:
The patent implements dynamic model selection and weighting where the system automatically adjusts which neural network architectures to use and what weights to assign to each based on current data conditions. During periods of high volatility or pandemic conditions, the system dynamically selects models that have demonstrated better performance, preventing overfitting by adapting to changing market conditions rather than relying on a fixed model structure.
Solution Approach 2:
The system incorporates feedback mechanisms where prediction errors are continuously monitored and used to adjust model weights and selection. The patent uses validation datasets and performance metrics to evaluate each neural network's predictions, then feeds this information back to adjust the ensemble weighting, ensuring that models generalize well to new conditions and preventing overfitting to historical patterns.
3Ease of operation
If statistical forecasting methods are used, then manual intervention is required in procurement and inventory management, but automated forecasting systems have high error rates
Solution Approach 1:
The patent implements a self-service forecasting system that automatically handles data preprocessing, model selection, training, and prediction without requiring manual intervention. The system autonomously manages the entire forecasting pipeline, from ingesting historical data to generating predictions, thereby eliminating the need for manual procurement and inventory management while achieving high accuracy through its ensemble neural network approach.
Solution Approach 2:
The forecasting system is designed to be universally applicable across multiple materials and market conditions. The patent creates a multi-functional platform that can handle diverse product types, pricing patterns, and demand fluctuations using the same neural network ensemble architecture, providing automated forecasting with high accuracy across different scenarios without requiring material-specific customization.
4Stability of the object's composition
If traditional forecasting models are used, then the system can handle continuous data, but it fails to accurately predict during sudden surge and volatile pricing and demand
Solution Approach 1:
The patent employs dynamic neural network models (LSTM and GRU) that are specifically designed to handle time-series data with abrupt changes and non-stationarity. These models can dynamically adapt to sudden surges and volatile conditions by learning temporal dependencies and adjusting to changing patterns, unlike traditional statistical models that assume data stationarity and fail during sudden market shifts.
Data Source
AI summary
Forecasting solutions including customization of raw data using uncertainty coefficient and using ensemble neural network architecture. The raw data is customized by cleaning and augmenting to obtain a processed dataset that is non-discreet and continuous, and; the ensemble neural network architecture is customized to include plurality of dependent and independent features to obtain ensembled weights from ensemble recurrent neural network (RNN) type architecture, a one versus rest ensemble RNN architecture and a forest of ensemble, to obtain an appropriate forecasting model that includes dynamically adaptive weights from ground truth along with the ensembled weights to create a final weighted output such that the final weighted output/forecast result has more accuracy and reduced false positives. The present invention allows for feedback mechanism from disruptive forecasting results if any from query module to go back to training module that enhances the accuracy of the next result without re-training.