Short-Term Electric Load Forecasting Using Echo State Networks and PCA Decomposition

Filippo Maria Bianchi, Enrico De Santis, Antonello Rizzi, Alireza SadeghianIEEE Access 2015 · OpenReview

Abstract

In this paper, we approach the problem of forecasting a time series (TS) of an electrical load measured on the Azienda Comunale Energia e Ambiente (ACEA) power grid, the company managing the electricity distribution in Rome, Italy, with an echo state network (ESN) considering two different leading times of 10 min and 1 day. We use a standard approach for predicting the load in the next 10 min, while, for a forecast horizon of one day, we represent the data with a high-dimensional multi-variate TS, where the number of variables is equivalent to the quantity of measurements registered in a day. Through the orthogonal transformation returned by PCA decomposition, we reduce the dimensionality of the TS to a lower number k of distinct variables; this allows us to cast the original prediction problem in k different one-step ahead predictions. The overall forecast can be effectively managed by k distinct prediction models, whose outputs are combined together to obtain the final result. We employ a genetic algorithm for tuning the parameters of the ESN and compare its prediction accuracy with a standard autoregressive integrated moving average model.

Introduction: The Challenge of Load Forecasting

Short-Term Electric Load Forecasting (STLF) is a critical component of modern power grid management. Accurate predictions are essential for balancing supply and demand; overestimation leads to wasted energy and higher costs, while underestimation can result in grid instability or the need for expensive emergency power purchases.

Traditional Recurrent Neural Networks (RNNs) are often used for time-series forecasting due to their ability to capture temporal dependencies. However, they face significant hurdles when applied to long-range horizons, such as a 24-hour forecast. In these scenarios, errors tend to compound over multiple time steps, and the network's "memory" of past states can fade. Additionally, tuning the hyperparameters of complex recurrent architectures is often a manual and inefficient process.

Methodology

To address these challenges, this research proposes a hybrid approach tailored to the forecasting horizon.

Reservoir Computing with Echo State Networks

For short-term (10-minute) forecasts, the study utilizes an Echo State Network (ESN). ESNs belong to the reservoir computing paradigm, where a large, fixed, random recurrent reservoir maps input signals into a high-dimensional state space. Unlike standard RNNs, only the output layer (the "readout") is trained using simple linear regression. This makes ESNs computationally efficient and avoids the vanishing gradient problems typical of backpropagation through time.

PCA Decomposition for Long-Horizon Forecasts

Predicting 24 hours ahead (144 steps at 10-minute intervals) directly is prone to error. Instead, the authors treat each day as a 144-dimensional vector and apply Principal Component Analysis (PCA). By reducing these 144 variables to just five principal components, which capture 94.6% of the total variance, the complex multi-step problem is transformed into five independent one-step-ahead prediction tasks.

Sample autocorrelation of the load data (a) and its first difference (b), illustrating the strong periodic dependencies that PCA helps manage.

Hyperparameter Optimization

To eliminate manual tuning, a Genetic Algorithm (GA) is employed to optimize the ESN’s internal parameters, including the spectral radius (controlling memory duration) and connectivity density. This ensures the reservoir is optimally configured for the specific dynamics of the electricity load.

Dataset and Analysis

The study utilizes real-world data from the ACEA power grid in Rome, Italy, specifically from the "Belsito Prisciano" Medium Voltage feeder. The dataset spans three years (2009–2011) with measurements taken every 10 minutes, totaling over 137,000 data points.

The electricity load profile over several days, demonstrating the clear periodic patterns and volatility inherent in grid demand.

Analysis of the data reveals a strong 24-hour seasonality. The researchers found that periodically updating the PCA rotation matrix with the most recent data significantly improved accuracy, as the underlying relationships between load variables can shift over time.

Results and Discussion

The performance of the ESN-PCA hybrid was compared against a standard ARIMA (Auto-Regressive Integrated Moving Average) model using Normalized Root Mean Squared Error (NRMSE).

Task ARIMA Error (NRMSE) ESN Error (NRMSE)
10-min Forecast 0.1429 0.1164
1-day Forecast (Direct) 0.5104 0.3254
1-day Forecast (PCA + Update) 0.3043 0.2524

The results demonstrate that the ESN consistently outperforms ARIMA. Most notably, the combination of PCA and periodic model updates for the 1-day forecast reduced the error by nearly 50% compared to the baseline ARIMA model.

Conclusion

This research highlights the power of combining reservoir computing with dimensionality reduction for complex time-series tasks. By predicting the "shape" of a daily load profile via PCA rather than individual future points, the model effectively bypasses the compounding error problem of long-horizon forecasting. While the Genetic Algorithm adds computational cost during the tuning phase, the resulting ESN models are highly efficient and accurate. Future work may explore integrating deep learning architectures or fuzzy logic systems to further enhance the readout layer's predictive capabilities.

Summary updated

Abstract

In this paper, we approach the problem of forecasting a time series (TS) of an electrical load measured on the Azienda Comunale Energia e Ambiente (ACEA) power grid, the company managing the electricity distribution in Rome, Italy, with an echo state network (ESN) considering two different leading times of 10 min and 1 day. We use a standard approach for predicting the load in the next 10 min, while, for a forecast horizon of one day, we represent the data with a high-dimensional multi-variate TS, where the number of variables is equivalent to the quantity of measurements registered in a day. Through the orthogonal transformation returned by PCA decomposition, we reduce the dimensionality of the TS to a lower number k of distinct variables; this allows us to cast the original prediction problem in k different one-step ahead predictions. The overall forecast can be effectively managed by k distinct prediction models, whose outputs are combined together to obtain the final result. We employ a genetic algorithm for tuning the parameters of the ESN and compare its prediction accuracy with a standard autoregressive integrated moving average model.
BibTeX
@inproceedings{bianchi2015short,
  author = {Filippo Maria Bianchi and Enrico De Santis and Antonello Rizzi and Alireza Sadeghian},
  title = {{Short-Term Electric Load Forecasting Using Echo State Networks and PCA Decomposition}},
  booktitle = {IEEE Access},
  year = {2015},
  url = {https://openreview.net/forum?id=07yoo0Uxjy1},
}
Figure