TL;DR
This paper introduces a novel transfer learning framework for smart grid load forecasting that significantly reduces reliance on local data, making it robust against cyber threats that could compromise data integrity. The method achieves improved prediction accuracy and enables more cost-effective energy scheduling by leveraging data from neighboring grids and accounting for temporal shifts in data distributions.
What Problem Does This Paper Address?
Smart grids, while offering enhanced efficiency and reliability, are vulnerable to cyber-attacks. A critical challenge arises when such attacks lead to missing or corrupted local grid data. This data loss cripples essential functions like load forecasting and power planning, which are crucial for grid operation. Traditional forecasting methods often require extensive local historical data, making them impractical when this data is unavailable. Furthermore, the inherent intermittency of renewable energy sources and the dynamic nature of power systems introduce significant variations in data distributions over time, posing a challenge for models trained on static historical data. This paper aims to address the problem of load forecasting with insufficient local data, particularly in the context of cyber-threats, and to develop a robust power scheduling strategy that accounts for these forecasting limitations.
Methods
The proposed framework is built upon three core components:
- Transfer Learning-Based Power Forecasting: This component leverages knowledge from a "source domain" (e.g., data from a neighboring, well-functioning grid) to improve load prediction in a "target domain" where local data is limited or compromised. The Maximum Mean Discrepancy (MMD) is used to measure the similarity between source and target domains, guiding the selection of an appropriate source domain to avoid negative transfer.
- Adaptive Time Series Prediction with Temporal Covariate Shift (TCS) Modeling: To handle the changing data distributions over time, the paper proposes modeling time series from a "covariate shift" perspective. This involves splitting the time series into segments with the most dissimilar probability distributions using a dynamic programming approach. The model is then trained to generalize across these diverse segments, enhancing its robustness. Long Short-Term Memory (LSTM) networks are employed as the core forecasting model due to their ability to capture long-term dependencies in time series data.
- Optimal Economic Power Scheduling with Shared Energy Storage: Building on the improved load forecasts, the paper introduces a day-ahead optimal economic power scheduling model. This model incorporates a novel concept of a "shared energy storage station" (ESS) that serves multiple users. The objective is to minimize the daily operating cost for the user community by intelligently managing the charging and discharging of the shared ESS, thereby reducing reliance on grid purchases during peak hours and utilizing renewable energy more effectively.
Datasets
The study utilizes field load data from three regions in Western Australia for case studies. These regions represent different consumption patterns: two industrial areas (Case 1 and Case 2) and one residential area (Case 3). The data covers the period from May 1, 2015, to July 1, 2021. For training and validation, 80% of the historical data was used, with the remaining 20% reserved for testing. Input features for forecasting include the number of days, holidays, weekdays, holiday length, season, population, and temperature.
Key Insights and Ideas
- Cyber-Threat Resilience: The core innovation is creating a load forecasting system that doesn't solely depend on local data, making it inherently more resilient to cyber-attacks that could corrupt or delete this data.
- Leveraging Neighboring Data: Transfer learning proves effective in bridging the gap caused by missing local data by intelligently utilizing information from similar, well-functioning neighboring grids.
- Handling Data Distribution Shifts: The TCS modeling approach is crucial for adapting to the dynamic nature of power system data, where distributions change over time due to factors like renewable energy integration. This is a significant improvement over static forecasting models.
- Shared Energy Storage for Cost Savings: The concept of a shared ESS offers a practical and economically viable solution for microgrids, reducing individual investment costs and optimizing overall energy management.
- Importance of Optimal Data Splitting: The paper highlights that the effectiveness of transfer learning is significantly influenced by how the time series data is segmented. Dividing data into segments with maximal distributional differences is key to achieving better generalization.
Key Results
- Improved Forecasting Accuracy: The proposed TCS-transfer learning method achieved a maximum improvement of 52.8% in Mean Absolute Percentage Error (MAPE) compared to other transfer learning-based methods and up to 64.4% compared to traditional methods.
- Reduced Operating Costs: The optimal economic power scheduling model using the shared ESS resulted in a reduction of the total operating cost by 36.1% compared to independent energy storage configurations.
- Addressing Energy Curtailment: The framework effectively addresses energy curtailment by better utilizing renewable energy sources and managing demand through the shared ESS.
- Faster Convergence (for some methods): While the proposed TCS-transfer learning method is slightly slower than the transformer in convergence speed, it offers significantly better prediction accuracy.
Limitations and Future Work
- Transformer Extension: The authors plan to explore a deeper extension of TCS-transfer learning to transformer architectures for potentially better performance.
- Centralized vs. Distributed ESS: The current work focuses on a centralized energy storage system. Future work will investigate optimal coordination for multiple distributed energy storage systems.
- Harmonics and Power Quality: A limitation of shared energy storage stations is their potential to cause harmonics, which can affect power quality. Further research is needed to address this.
- Data Availability: While the paper discusses data sources, some specific net power data used for findings is under embargo, with requests to be considered by the corresponding author after publication.
Summary updated