Research on Small Hydropower Generation Forecasting Method Based on Improved BP Neural Network

Miao Li, Chang Deng, Jin Tan, Wei Yang, Li Zheng2016 · DOI

Abstract

Small hydro-power generation shows strong uncertainty, which greatly affects the load forecasting work in small hydropower regions. Thus it's important to improve the accuracy of small hydropower generation load forecasting. At present, the most commonly used forecasting method is artificial neural network, which has strong adaptability and learning ability but poor generalization and easily falls into local minimum. The random fluctuation of small hydropower is not taken into consideration. This paper was based on analyzing the characteristics of small hydropower generation load, combining the wavelet transform to decompose the historical load to establish the prediction model for each component feature. Particle Swarm Optimization (Algorithm) was used to optimize initial weights and thresholds of neural networks before the prediction. After verified by real case in a rich small hydropower area in some province, the load prediction precision reaches 93.7%, higher than the precision of the high-voltage system criteria for assessing. The accuracy and effectiveness of the method is verified.

Introduction

As power grids transition into Active Distribution Networks (ADNs), they increasingly rely on Distributed Generation (DG) such as solar, wind, and small hydropower (SHP). While SHP is an environmentally friendly energy source, it is notoriously difficult to predict. Unlike large-scale dams, small hydropower is highly sensitive to local weather and watershed dynamics, exhibiting "strong uncertainty" and random fluctuations.

Traditional load forecasting methods often fail to account for these unique random fluctuations. Furthermore, standard Artificial Neural Networks (ANNs) frequently converge on "local minima"—sub-optimal solutions—during the training phase. This research addresses these challenges by proposing a hybrid model that integrates multi-resolution analysis with optimized machine learning.

Methodology

The proposed forecasting framework utilizes a "divide and conquer" strategy consisting of three primary stages:

1. Wavelet Decomposition

To manage the inherent volatility of SHP signals, the original historical load is decomposed into different frequency components using a db2 wavelet. This process separates the "low-frequency" trend (labeled M3), which represents the general base load, from "high-frequency" fluctuations (labeled N1, N2, and N3), which capture noisy, random variations.

Wavelet decomposition of the original load into trend (M3) and fluctuation components (N1, N2, N3).

2. Feature Selection and Cumulative Rainfall

Input variables are tailored for each decomposed component. In addition to historical load and temperature data, the model incorporates the cumulative effect of rainfall. By including rainfall data from the previous 3, 5, 7, and 10 days, the model captures the delayed hydrological response of the watershed, which is a critical driver of SHP generation.

3. PSO-BP Neural Network

To improve the generalization of standard Backpropagation (BP) networks, the authors employ Particle Swarm Optimization (PSO). PSO is used to determine the optimal initial weights and thresholds for the neural network. This optimization step ensures that the training process begins from a globally competitive position, significantly reducing the likelihood of the model getting stuck in local minima.

Results and Discussion

The model was validated using real-world historical data from a province in China characterized by abundant small hydropower resources. The dataset included 24-point daily load values, temperature records, and rainfall data.

The performance of the proposed hybrid model (Predicted Value 1) was compared against a standard forecasting model that does not utilize wavelet decomposition (Predicted Value 2). As shown in the results, the hybrid model follows the actual load curve much more closely, particularly during volatile peaks and troughs.

Comparison of actual load values with predictions from the proposed hybrid model (Value 1) and a standard model (Value 2).

The proposed method achieved a prediction precision of 93.7%, which exceeds the precision requirements of standard high-voltage system criteria for assessment. This demonstrates that predicting "noisy" high-frequency components separately and then reconstructing the signal yields significantly higher accuracy than attempting to predict the raw, volatile signal directly.

Conclusion

This research demonstrates that a hybrid approach combining Wavelet Decomposition and PSO-optimized BP Neural Networks can effectively handle the uncertainty of small hydropower generation. By accounting for the cumulative effects of rainfall and decomposing complex signals into manageable components, the model provides a reliable tool for load forecasting in modern active distribution networks. Future work may involve exploring different wavelet types and integrating social electricity load data to further refine regional forecasting.

Summary updated

Abstract

Small hydro-power generation shows strong uncertainty, which greatly affects the load forecasting work in small hydropower regions. Thus it's important to improve the accuracy of small hydropower generation load forecasting. At present, the most commonly used forecasting method is artificial neural network, which has strong adaptability and learning ability but poor generalization and easily falls into local minimum. The random fluctuation of small hydropower is not taken into consideration. This paper was based on analyzing the characteristics of small hydropower generation load, combining the wavelet transform to decompose the historical load to establish the prediction model for each component feature. Particle Swarm Optimization (Algorithm) was used to optimize initial weights and thresholds of neural networks before the prediction. After verified by real case in a rich small hydropower area in some province, the load prediction precision reaches 93.7%, higher than the precision of the high-voltage system criteria for assessing. The accuracy and effectiveness of the method is verified.
BibTeX
@misc{li2016research,
  author = {Miao Li and Chang Deng and Jin Tan and Wei Yang and Li Zheng},
  title = {{Research on Small Hydropower Generation Forecasting Method Based on Improved BP Neural Network}},
  year = {2016},
  doi = {10.2991/icmemtc-16.2016.214},
}
Figure