TL;DR
This paper introduces a novel ultra-short-term photovoltaic (PV) power prediction model called LERENN, which integrates chaos theory and neuropsychological emotion models. By reconstructing PV power time series in a chaotic phase space and incorporating "emotional" parameters, LERENN aims to improve prediction accuracy, especially during rapid power fluctuations and ramp events, outperforming traditional models like LSTM and BPNN.
What Problem Does This Paper Address?
Predicting photovoltaic (PV) power output is crucial for grid stability, but it's challenging due to the inherent intermittency and randomness of solar generation, influenced by weather. Existing data-driven models struggle to achieve high accuracy, particularly for ultra-short-term (seconds to an hour) predictions. This is exacerbated by the difficulty in capturing the complex, non-linear dynamics of PV power and the sudden, rapid changes (ramp events) that occur under varying weather conditions. The paper aims to bridge this gap by developing a model that can better capture these hidden dynamics and react more sensitively to abrupt changes.
Methods
The core of the LERENN model lies in a multi-stage approach:
- Chaos Theory and Phase Space Reconstruction: The paper first applies chaos theory to PV power time series, assuming underlying deterministic laws within the apparent randomness. It uses phase space reconstruction techniques (specifically the C-C method) to transform the 1D time series into a higher-dimensional space, revealing hidden characteristics and dynamics. This process involves determining optimal embedding dimensions and delay times.
- Single-Step Cyclic Rolling Localized Prediction: A localized prediction mechanism is employed, where the model focuses on predicting the immediate future based on the most relevant recent data points. This "rolling" approach updates the prediction window with each new data point, aiming to reduce error accumulation.
- Localized Emotion Reconstruction (LERENN): This is the novel aspect. Inspired by neuropsychology, the model incorporates "emotional" parameters (anxiety and confidence coefficients) within a limbic-based artificial emotional neural network (LiAENN) architecture. These parameters are dynamically adjusted based on prediction errors and the learning process, mimicking how humans focus attention on new or "interesting" data. The "expanded signal," derived from the reconstructed phase space, also plays a role in this emotional processing. The model aims to make the prediction more sensitive to the specific characteristics of the PV power time series.
Datasets
The study utilizes PV power data from a grid-connected PV power station at the National Institute of Standards and Technology (NIST) in Gaithersburg, MD.
- Training Data: 62 days of historical PV power data.
- Forecasting Data: 8 days of data, specifically chosen to represent different weather conditions: 2 sunny, 2 cloudy, 2 overcast, and 2 abrupt weather days (sunny to cloudy, cloudy to sunny).
- The data resolution is 5-minute intervals.
Key Insights and Ideas
- Chaos Theory for PV Power: The paper posits that PV power time series, despite appearing random, exhibit chaotic characteristics that can be exploited for better prediction. Reconstructing the phase space helps uncover these underlying deterministic patterns.
- "Emotional" Intelligence in Prediction: The integration of emotional parameters (anxiety and confidence) into the neural network is a novel approach. The idea is that by simulating an "attentional" mechanism, the model can better focus on crucial data points and adapt more intelligently to changing conditions, akin to human learning.
- Localized Prediction for Robustness: The cyclic, localized prediction mechanism is key to avoiding error propagation and ensuring that each prediction step is relatively independent of previous prediction errors.
- Sensitivity to Ramp Events: The combination of chaos theory and emotional parameters appears to make the LERENN model more sensitive to sudden changes in PV power output, which is a critical challenge in ultra-short-term forecasting.
Key Results
The LERENN model was compared against Persistence Model (PM), Long Short-Term Memory (LSTM), Limbic-based Artificial Emotional Neural Network (LiAENN), and Back Propagation Neural Network (BPNN).
- Single-Step Prediction: LERENN generally outperformed the benchmark models across various weather conditions, showing lower MAPE, RMSE, and SSE. It was particularly effective in detecting ramp events.
- Three-Step Prediction: While all models' accuracy decreased with longer prediction horizons, LERENN showed better robustness, with its performance deteriorating less significantly than other models.
- Weather Adaptability: The model demonstrated good adaptability across different weather conditions, including challenging scenarios like abrupt weather changes and cloudy days.
- Error Reduction: LERENN was effective in reducing large prediction errors, especially in the first error bin (smallest relative error), indicating improved accuracy for a significant portion of predictions.
Limitations and Future Work
- Five-Minute Resolution Applicability: The authors acknowledge that a five-minute prediction resolution might be less directly applicable for immediate power quality smoothing in real-world utility operations, suggesting the need for further consideration of other metrics.
- Further Improvement: While LERENN shows improved accuracy, there's still room for enhancement, particularly on overcast days where prediction accuracy is limited.
- Future Research Directions:
- Incorporating meteorological factors like solar radiation intensity and aerosol index as additional inputs to further refine predictions.
- Combining prediction results with energy storage systems for more comprehensive grid-connected PV generation system power output smoothing.
- Exploring other metrics beyond point-wise accuracy to provide a more complete picture of prediction performance.
Summary updated