Abstract
In order to improve the forecast accuracy and stability of the micro-grid uncontrollable remaining load and provide a more reliable basis for micro-grid power generation plan, an ultra-short-term micro-grid uncontrollable remaining load forecasting model based on the improved online sequential Extreme Learning Machine is proposed. Aimed at the wind and solar power generation and load characteristics, the weight update of old and new training data is added to the Extreme Learning Machine. And the average value of multi-module is used to enhance the predict stability of the algorithm. After the real data from UCI Machine Learning Repository is analyzed, the result shows that the algorithm is superior to the traditional Extreme Learning Machine (ELM) and the online sequential Extreme Learning Machine (OS-ELM) and the proposed algorithm is feasible.
What Problem Does This Paper Address?
Micro-grids are localized energy grids that can operate independently, often relying on "uncontrollable" renewable sources like wind turbines and solar panels. Because wind and sun are intermittent, grid operators need to know the remaining load: the total electricity demand minus the power generated by these uncontrollable sources.
Accurately forecasting this remaining load is essential for planning the output of "controllable" sources (like diesel generators or batteries) to ensure grid stability. Traditional methods like Support Vector Machines (SVM) or standard Neural Networks often suffer from slow training speeds or poor generalization. While the Extreme Learning Machine (ELM) is much faster, it struggles with the streaming nature of real-world data and can be unstable, leading to "prediction accidents" where a single model produces a wildly inaccurate result.
Methods
The authors propose a three-layered improvement over the standard Extreme Learning Machine, resulting in the MM-WOS-ELM (Multi-Module Weighted Online Sequential Extreme Learning Machine) architecture:
- Online Sequential Learning (OS-ELM): Instead of retraining the entire model when new data arrives, OS-ELM updates the model using only the latest block of data, saving massive amounts of time and computational resources.
- Weighting (WOS-ELM): Not all data blocks are equally useful. The authors introduced a weight factor ($\alpha$) for new data. If the model’s recent error (RMSE) is high, the weight of the new data is adjusted (either 0.995 or 1.005) to help the model "refocus" and improve accuracy.
- Multi-Module Ensemble (MM-WOS-ELM): To prevent the instability inherent in a single OS-ELM run, the authors deploy multiple modules ($M$) simultaneously. The final forecast is the average of all these modules, which smooths out outliers and provides a more reliable prediction.
Datasets
The study utilized real-world data from the UCI Machine Learning Repository, spanning from March 13, 2012, to April 4, 2012.
- Load Data: Included timestamps and temperature.
- Environmental Data: Included wind speed, light intensity, temperature, precipitation, and atmospheric humidity.
- Preprocessing: All datasets were normalized before being fed into the models to ensure features were on a comparable scale.
Key Insights and Ideas
- The "Remaining Load" Strategy: Rather than just predicting demand, the authors focus on the net demand. This is the most actionable metric for a micro-grid controller.
- The Stability-Accuracy Tradeoff: The authors found that while OS-ELM is fast, its randomness can lead to instability. Their ensemble approach mitigates this randomness without sacrificing the speed benefits of ELM.
- Optimal Complexity: Through testing, the researchers discovered that more modules aren't always better. They used a "comprehensive index" (Syn) that balances accuracy, stability, and computation time, finding that 5 modules provided the optimal performance for this specific application.
Key Results
The proposed MM-WOS-ELM outperformed traditional ELM, OS-ELM, and the single-weighted WOS-ELM across all metrics:
- Accuracy (RMSE): The MM-WOS-ELM achieved an RMSE of 2.2956, compared to 2.7976 for standard ELM and 2.4817 for OS-ELM.
- Stability (Standard Deviation of RMSE): This is where the model truly shined. The stability measure for the proposed model was 0.0435, nearly three times more stable than the standard ELM (0.1324).
- Feasibility: The model successfully tracked the actual uncontrollable remaining load profile, demonstrating its readiness for ultra-short-term forecasting.
Limitations and Future Work
While the paper demonstrates a clear improvement in stability, it relies on somewhat heuristic values for the weight adjustment ($\alpha = 0.995$ or $1.005$). Future research could explore more dynamic or mathematically optimized ways to determine these weights. Additionally, the authors suggest integrating these high-accuracy forecasts into an economic dispatch plan to optimize the cost of filling the energy gap with controllable power sources.
Summary updated