Abstract
Accurate reactive load prediction can improve the accuracy and process of reactive power optimization for power grids and improve the control effect. The changes in the bus reactive load and active load are not synchronous, the base of the reactive load is small, nonlinear changes are abundant, and it is difficult to mine the inherent data trends. In view of the above problems, this paper proposes a method for predicting bus reactive loads based on deep learning. A bus reactive load prediction model is constructed based on a dual-input long short-term memory neural network to mine the detailed characteristics of active and reactive load data. Active and reactive loads are used as input and output data for the dynamic modeling of load time series data to form integrated forecasts of bus active and reactive loads. The experimental results show that this method can accurately predict the reactive power load of buses, and the prediction accuracy is better than that of time series and general long short-term memory neural network prediction models.
Introduction and Problem Statement
In modern power systems, maintaining a balance of reactive power is essential for voltage stability and grid efficiency. However, predicting reactive load (measured in MVar) is notoriously difficult compared to active load (MW). The authors identify several key challenges that hinder accurate forecasting:
- Non-synchronicity: Changes in reactive load do not perfectly mirror active load, making simple scaling models ineffective.
- Small Base Values: Reactive loads often have smaller magnitudes, making them more sensitive to "noise" and fluctuations.
- High Nonlinearity: Reactive power is influenced by complex factors like substation compensation, line flow, and diverse consumer behavior, leading to irregular data trends.
- Lagging Control: Current grid management often relies on post-correction. Accurate prediction is the first step toward proactive, fine-scale voltage control.
Methodology: The Dual-Input LSTM
The researchers propose a Dual-Input LSTM (DI-LSTM) architecture. While standard LSTMs are effective for time-series data, the DI-LSTM specifically exploits the relationship between different types of power data to improve accuracy.
- Architecture: The model uses a stacked structure consisting of two LSTM layers to capture long-term temporal dependencies, followed by a single Deep Neural Network (DNN) layer with 96 hidden units for final output mapping.
- Dual Inputs: The model takes historical active load, historical reactive load, and a "day type" variable (indicating workdays vs. holidays) as inputs. This allows the model to learn the complex, non-linear coupling (and decoupling) between the two power types.
- Data Processing: To handle the different scales of active and reactive power, min-max normalization is applied. The model uses 96 sampling points per day (15-minute intervals) to predict the load for the following day.
- Optimization: The network is trained using the Adam optimizer and Backpropagation Through Time (BPTT) to mitigate the vanishing gradient problem.
Experimental Dataset
The study utilizes real-world SCADA (Supervisory Control and Data Acquisition) data from a regional power grid in Hainan Province, China. The dataset includes 357 bus loads sampled at 15-minute intervals from June 1 to August 22, 2011. The training set covers data up to August 5, while the remainder is used for testing.
Key Results and Insights
The DI-LSTM was compared against traditional ARMA (Autoregressive Moving Average) models and standard single-input LSTMs. The results demonstrate the clear superiority of the proposed deep learning approach:
- Reactive Power Accuracy: At the "Jinjiang Station 2-bus," the ARMA model failed significantly with a Mean Absolute Percentage Error (MAPE) of 240.049%. The standard LSTM achieved 32.026%, while the proposed DI-LSTM slashed the error to just 10.148%.
- Active Power Accuracy: Even for active power, the DI-LSTM showed superior performance with a MAPE of ~1.1% compared to ~1.8% for the standard LSTM.
- Linear Model Failure: The paper highlights that traditional statistical models like ARMA fail during significant reactive load shifts because they rely on linear fitting, which cannot account for the inherent volatility of reactive power.
- Day-Type Correction: Including a holiday/workday variable acts as a vital correction factor, accounting for drastic changes in human behavior and power consumption patterns.
Conclusion and Future Work
The DI-LSTM model provides a robust solution for predicting highly non-linear reactive loads, enabling more proactive grid management. While successful, future work could incorporate external factors such as weather data (temperature and humidity) and test the model's generalizability across grids with different industrial and residential ratios. Integrating these predictions into real-time automated reactive power optimization hardware remains a key goal for future deployment.
Summary updated