Abstract
Forecasting complex, non-linear systems remains a significant challenge in data science. Traditional single-model approaches often fail to capture the full spectrum of data volatility, while static combination models lack the adaptability required for real-time shifts. This paper introduces an integrated parallel forecasting framework that synergizes modified Fuzzy Time Series (FTS) and Support Vector Machines (SVM). By employing Improved Fuzzy C-Means (IFCM) clustering for data preprocessing and a dynamic weighting scheme for output integration, the proposed model achieves a significant reduction in predictive error compared to conventional methods.
Introduction
Predicting trends in complex systems—such as financial markets or equipment management—is difficult because data often exhibits non-linear, non-stationary characteristics. Single models like Support Vector Machines (SVM) are powerful for small-sample regression but are sensitive to parameter selection and data distribution. Conversely, Fuzzy Time Series (FTS) models offer flexibility but frequently rely on arbitrary, fixed-length intervals that ignore the underlying data density.
Existing hybrid models typically follow a serial or static parallel structure. Serial models risk error propagation, where inaccuracies from the first stage are amplified in the second. Static parallel models use fixed weights (e.g., a simple average), failing to adjust when one sub-model's performance degrades. This research addresses these gaps by creating a parallel system where both models are individually optimized and then combined using time-varying, dynamic weights.
Proposed Methodology
The core of the proposed system is a parallel architecture that processes input data through two distinct, optimized paths before merging the results based on historical performance.
1. Improved Fuzzy C-Means (IFCM) Clustering
To enhance the quality of the training data, the authors modified the standard Fuzzy C-Means (FCM) algorithm. The IFCM approach better handles outliers and maximizes inter-cluster distance while minimizing intra-cluster variance. This ensures that the subsequent SVM training focuses on the most representative data neighborhoods, effectively denoising the input set.
2. Non-Isometric Multi-Scale FTS
Instead of traditional equal-sized intervals, this model utilizes a "ratio-based" domain partition. By using an echelon interval system to determine membership degrees, the FTS model becomes sensitive to the actual density of data points, capturing micro-trends that standard models often smooth over.
3. MPSO-SVM Optimization
To solve the parameter-tuning problem in SVM, a Modified Particle Swarm Optimization (MPSO) algorithm is used. It introduces particle diversity and an exponential-decline inertia weight to prevent the search from getting trapped in local optima, ensuring the discovery of the global best parameters for the forecasting task.
4. Dynamic Weighting Scheme
The final output is not a simple average. The system treats the combination weights as a time series themselves, calculating them based on the recent accuracy of each sub-model. A fuzzy model is then used to forecast the optimal weights for the next time step, allowing the system to "lean" into whichever model is currently performing better.
Experimental Results
The model was rigorously tested using real-world Motor Hour Profit and Loss data from 2002 to 2006. This dataset is characterized by high volatility, providing a robust benchmark for forecasting precision.
| Method | RMSE (Root Mean Square Error) | MAPE (Mean Absolute Percentage Error) |
|---|---|---|
| Improved Parallel Model (Proposed) | 11.66 | 7.92% |
| Parallel Model (Standard) | 14.37 | 12.74% |
| Serial Model | 15.01 | 12.89% |
| Improved FTS | 14.63 | 12.41% |
| MPSO-SVM | 15.84 | 13.45% |
The results demonstrate that the proposed integrated model reduces the Root Mean Square Error (RMSE) by approximately 19-22% compared to standard parallel or serial combination models.
Conclusion
By integrating IFCM clustering, ratio-based FTS partitioning, and MPSO-optimized SVM within a dynamic parallel framework, this research provides a highly adaptive solution for non-linear forecasting. The dynamic weighting mechanism is particularly effective, as it allows the system to self-correct based on the evolving accuracy of its sub-models. Future work will focus on reducing the computational overhead of the iterative optimization processes to make the model more scalable for larger datasets.
Summary updated