TL;DR
This paper introduces a novel two-stage framework for short-term electricity load forecasting that combines a Residual Convolutional Neural Network (R-CNN) for feature extraction with a Multilayer Long Short-Term Memory (ML-LSTM) network for temporal sequence learning. The proposed R-CNN-ML-LSTM model demonstrates superior accuracy in predicting electricity consumption compared to existing state-of-the-art methods on both residential and commercial datasets.
What Problem Does This Paper Address?
Accurate electricity load forecasting is crucial for efficient smart grid management, enabling better energy allocation, peak load regulation, and ultimately, customer financial benefits. However, predicting electricity consumption is a complex task due to its inherent nonlinearity and dependence on numerous factors like time, weather, and economic conditions. Existing machine learning models, while powerful, often struggle to effectively capture both the spatial patterns and temporal dependencies present in electricity consumption data, leading to suboptimal forecasting accuracy. This paper aims to address this challenge by developing a more robust and efficient forecasting model.
Methods
The proposed approach employs a two-stage framework:
- Data Preprocessing: This initial stage focuses on cleaning and preparing the raw electricity consumption data. It involves:
- Outlier Removal: Utilizing the "three sigma rules" to identify and remove anomalous data points.
- Missing Value Imputation: Employing a recovering interpolation method to fill in gaps in the data.
- Data Transformation: Applying the Box-Cox power transformation to make the data more Gaussian-like, improving its suitability for subsequent modeling.
- Normalization: Using the Min-Max scaler to bring the data into a specific range, which is essential for deep learning networks sensitive to data scale.
- R-CNN with ML-LSTM Architecture: This is the core of the proposed model.
- R-CNN (Residual Convolutional Neural Network): This component is responsible for extracting spatial features and patterns from the preprocessed electricity consumption data. R-CNNs are known for their ability to learn hierarchical features and overcome the degradation problem encountered in very deep conventional CNNs.
- ML-LSTM (Multilayer Long Short-Term Memory): The features extracted by the R-CNN are then fed into the ML-LSTM network. LSTMs are well-suited for learning sequential information and temporal dependencies, which are critical for time-series forecasting. The multilayer aspect allows for learning more complex temporal relationships.
The R-CNN extracts spatial features, and these features are then passed to the ML-LSTM for temporal learning, creating a hybrid model that leverages the strengths of both architectures.
Datasets
The proposed model was evaluated on two benchmark datasets:
- IHEPC Dataset: This dataset was collected from a residential building in France between 2006 and 2010. It includes minute-level data with attributes such as date and time, global active/reactive power, intensity, average ambient temperature, and voltage. For daily load forecasting, the minute-level data was down-sampled to hourly resolutions.
- PJM Dataset: This dataset represents electricity load forecasting data from the Pennsylvania-New Jersey-Maryland (PJM) region, which operates the eastern electricity grid of the US. The data is recorded in one-hour resolution in megawatts.
Key Insights and Ideas
- Synergy of R-CNN and ML-LSTM: The paper highlights the effectiveness of combining R-CNN for spatial feature extraction with ML-LSTM for temporal sequence learning. This hybrid approach addresses the limitations of individual models by capturing both the intricate patterns within the data and their evolution over time.
- Importance of Data Preprocessing: The study emphasizes that robust data preprocessing, including outlier removal and missing value imputation, is crucial for achieving accurate forecasting results. The chosen methods (three sigma rules, interpolation, Box-Cox, Min-Max scaling) are shown to be effective in preparing the data for the deep learning model.
- Addressing Nonlinearity: Electricity consumption is inherently nonlinear. The R-CNN's ability to learn complex, hierarchical features and the LSTM's capacity to model sequential dependencies are key to handling this nonlinearity effectively.
- Improved Performance on Diverse Datasets: The model demonstrates strong performance on both residential (IHEPC) and commercial (PJM) datasets, suggesting its generalizability across different consumption patterns.
Key Results
The proposed R-CNN with ML-LSTM model achieved significantly lower error rates compared to various state-of-the-art baseline models on both the IHEPC and PJM datasets.
- IHEPC Dataset (Hourly): Achieved error rates of RMSE: 0.0325, MAE: 0.0144, MSE: 0.0011, R²: 0.9841, and MAPE: 1.024.
- IHEPC Dataset (Daily): Achieved error rates of RMSE: 0.0447, MAE: 0.0132, MSE: 0.002, R²: 0.9759, and MAPE: 2.457.
- PJM Dataset: The model consistently achieved the lowest error rates across 10 different regions of the PJM dataset, with specific examples showing very low MAPE values (e.g., 0.5504 for AEP, 0.4982 for DAYTON, 0.2864 for PJMW).
Limitations and Future Work
- Limitations: The paper focuses on short-term electricity load forecasting. The performance on medium-term and long-term forecasting is not explored.
- Future Work: The authors suggest several avenues for future research:
- Extending the model to medium-term and long-term electricity load forecasting.
- Integrating environmental sensor data (e.g., weather, humidity) to further enhance prediction accuracy.
- Investigating the model's performance in other prediction domains, such as fault prediction, renewable power generation prediction, and traffic flow prediction.
Summary updated