PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Photovoltaic Nowcasting

Jinsong Zhang, Rodrigo Verschae, Shohei Nobuhara, Jean-François Lalonde

arXiv:1810.06327v1cs.CV

TL;DR

Minute-scale photovoltaic nowcasting is important for smart-grid management, yet cloud appearance and motion are difficult to model explicitly from camera images. The paper learns the relationship between historical power, sky images, and future output with MLP, CNN, and LSTM architectures. On the 1-minute task, the LSTM achieves a 21% RMSE skill score over persistence and outperforms the alternatives.

  • Problem

    Minute-scale photovoltaic forecasting is important for smart-grid management, but cloud dynamics make explicit computer-vision modeling from sky images challenging.

  • Method

    The paper trains and compares MLP, CNN, and LSTM deep-learning models that combine historical photovoltaic values with nearby sky images to predict future output.

  • Results

    21% RMSE skill score over the persistence baseline is achieved by the LSTM model on the 1-minute photovoltaic forecasting task.

  • Takeaways & Limitations

    The experiments identify temporal modeling with LSTM, auxiliary learning sub-tasks, and sky images as important aspects of the reported forecasting performance.

  • Takeaways & Limitations

    The study relies on data from a single site, camera, and photovoltaic panel, limiting demonstrated cross-site generality.

Abstract

from arXiv · show

Predicting the short-term power output of a photovoltaic panel is an important task for the efficient management of smart grids. Short-term forecasting at the minute scale, also known as nowcasting, can benefit from sky images captured by regular cameras and installed close to the solar panel. However, estimating the weather conditions from these images---sun intensity, cloud appearance and movement, etc.---is a very challenging task that the community has yet to solve with traditional computer vision techniques. In this work, we propose to learn the relationship between sky appearance and the future photovoltaic power output using deep learning. We train several variants of convolutional neural networks which take historical photovoltaic power values and sky images as input and estimate photovoltaic power in a very short term future. In particular, we compare three different architectures based on: a multi-layer perceptron (MLP), a convolutional neural network (CNN), and a long short term memory (LSTM) module. We evaluate our approach quantitatively on a dataset of photovoltaic power values and corresponding images gathered in Kyoto, Japan. Our experiments reveal that the MLP network, already used similarly in previous work, achieves an RMSE skill score of 7% over the commonly-used persistence baseline on the 1-minute future photovoltaic power prediction task. Our CNN-based network improves upon this with a 12% skill score. In contrast, our LSTM-based model, which can learn the temporal dependencies in the data, achieves a 21% RMSE skill score, thus outperforming all other approaches.

1. Introduction

Minute-scale photovoltaic nowcasting is needed for smart-grid management, but conventional weather data and explicit cloud modeling are limited. The paper therefore learns power-output relationships from historical panel values and nearby sky images using several deep-learning architectures.

  • Minute-scale photovoltaic nowcasting supports smart-grid operations including system integration, power continuity, and ramp-rate management.
  • NWP and satellite data are often unsuitable for nowcasting because their spatial and temporal resolution is limited, while higher-resolution alternatives can be expensive.
  • Regular cameras installed near panels provide inexpensive, high-resolution local observations, but their sky images require analysis to relate appearance to photovoltaic output.
  • Cloud variations in shape, appearance, velocity, and direction make explicit computer-vision modeling of future cloud appearance challenging.
  • The paper learns this relationship directly with deep networks combining past power outputs and images, and compares MLP, CNN, and LSTM architectures.
  • 21% RMSE skill score over persistence is achieved by the LSTM models on 1-minute photovoltaic forecasting.

2. Related work

Prior solar-forecasting methods use physical or data-driven approaches, often relying on weather or satellite data. This work instead learns how sky appearance and cloud motion translate into photovoltaic production for 1-minute forecasting.

  • Approaches: Physical solar-forecasting models depend heavily on numerical weather predictions, which are generally unsuitable for very short-term forecasting.
  • Data: NWP and satellite images provide regional weather information but typically lack adequate temporal resolution for very short-term forecasts.
  • Short-term forecasting: Prior work found that weather data may help over several-hour horizons, whereas short-term photovoltaic energy can be predicted from historical energy values alone.
  • Sky-image forecasting: Sky-image methods commonly estimate cloud cover and movement with image-processing techniques such as RGB ratios.
  • Novelty: This paper avoids explicit image processing and automatically learns how cloud motion and sky appearance translate into future photovoltaic production.
  • Deep learning: Deep-learning solar-forecasting studies have addressed hourly or day-ahead horizons, while this work targets the challenging 1-minute horizon.

3. Data

The dataset combines synchronized photovoltaic measurements with hemispherical sky images captured near a Kyoto University panel. Images use multiple exposures and are processed into minute-scale inputs for forecasting.

  • Data capture: The dataset uses a 2500W photovoltaic station with a 10 × 6 m2 panel installed at Kyoto University in Japan.
  • Data capture: A fisheye camera captured 1280×1280 sky images from approximately 180 m away, pointing toward the zenith.
  • Data capture: Four exposure times capture greater dynamic range, although the sun remains over-exposed even at the fastest exposure.
  • Data collection: Photovoltaic and image captures were synchronized, recorded over 1.5 years, and reduced to 90 randomly sampled days spanning different seasons.
  • Preprocessing: The experiments retain one-minute data by averaging photovoltaic values over each preceding 60-second interval and removing invalid images and zero-output records.
  • Image inputs: Each image input stacks four grayscale exposures from five images captured at 15-second intervals over the preceding minute.

4. Overview and notation

The paper formulates nowcasting as estimating a future photovoltaic power change from historical power values and, when available, current and previous sky images. The predicted change is then used to obtain the future power value.

  • Notation: Nowcasting predicts photovoltaic power at a future time typically about 1 minute after the current time.
  • Prediction target: The learning target is the estimated power variation, from which the future photovoltaic value is obtained.
  • Power-only inputs: The MLP-style formulation maps current and previous photovoltaic outputs to the predicted variation using trainable nonlinear network functions.
  • Image and power inputs: With image data, the model maps current and previous sky images together with power outputs to the predicted photovoltaic power variation.
  • Modeling approaches: The following architectures model this formulation using different deep-learning approaches.

5. Deep learning architectures to predict photovoltaic production

The paper develops three deep-learning architectures that combine historical photovoltaic power and sky-image information to forecast near-term power variation. The progression from an MLP baseline to CNN and LSTM designs adds image encoding and temporal modeling for this task.

  • 5.1. MLP with past photovoltaic values: The MLP baseline uses only historical photovoltaic values to forecast the future power variation.It has two hidden layers and one output neuron, with m = n = 64 selected as a compromise between performance and generalization.
  • 5.2. CNN integrating sky images: The CNN model combines encoded historical power values with independently encoded sky images before predicting power variation.A power MLP produces a 64-dimensional vector, while image encoders produce latent vectors that are concatenated and passed to a three-layer predictor.
  • 5.2. CNN integrating sky images: The image encoder compresses each sky image into a latent vector using convolutional layers, fire modules, normalization, ReLU activation, pooling, and residual links.Its final convolution layer maps the two-dimensional activation map into a one-dimensional vector for downstream prediction.
  • 5.3. LSTM for modeling temporal information: The LSTM architecture processes image latent vectors sequentially so its final output encodes temporal information from the historical image sequence.That output is concatenated with the historical-power representation and fed to a predictor for future power variation.
  • 5.4. Multi-task learning: The multi-task LSTM adds auxiliary regressors and predictors for absolute power, sun position, image reconstruction, sun-position variation, and sky-intensity variation.Instant and spatial tasks use the individual image latent vectors, whereas temporal tasks use the single LSTM output vector.

6. Experiments

The experiments compare deep learning architectures and preprocessing choices for 1-minute photovoltaic forecasting across weather conditions and forecast horizons. LSTM-based models generally outperform simpler alternatives, while performance declines at longer horizons.

  • Model comparison: All proposed models outperform the persistence baseline across SS-MAE and SS-RMSE, with skill scores computed separately for clear, partly cloudy, and overcast conditions.Higher skill scores indicate better performance relative to persistence.
  • Model comparison: CNN skill scores exceed those of the MLP model in all three weather conditions, showing that incorporating sky images improves prediction.The largest CNN gain occurs under partly cloudy conditions.
  • Model comparison: The LSTM achieves SS-MAE of 34.5, 25.5, 21.8 and SS-RMSE of 15.0, 22.3, 12.8 on clear, cloudy, and overcast conditions, respectively.These results are reported for the LSTM model across the three weather categories.
  • Multi-task learning: Adding regressors and predictors to form LSTM-Full produces performance similar to LSTM on SS-MAE but better balances results across weather conditions.Subsequent experiments use LSTM-Full because it is reported as the best-performing model.
  • Impact of different exposures: Using four exposures as separate image channels improves photovoltaic forecasting because different exposures capture different sky components, including the sun and clouds.The shortest exposure captures the brightest sun and surroundings, whereas the longest makes clouds visible.
  • Varying the time horizons: For clear skies, SS-MAE and SS-RMSE decline from 16.4% and 11.5% at 2 minutes to 12.1% and 7.7% at 10 minutes.The paper attributes declining skill at longer horizons partly to dramatic changes in sky appearance and suggests combining sky images with NWP or satellite images.

7. Discussion of limitations

The work is limited by its narrow data setting and difficulty forecasting abrupt photovoltaic changes. The authors discuss possible adaptation strategies and input improvements.

  • The dataset comes from one site, one camera, and one photovoltaic panel, limiting demonstrated generality.The authors suggest fine-tuning pretrained networks on another dataset as a potential remedy.
  • The LSTM and LSTM-Full models show little difference in overall performance.LSTM-Full is better-balanced across weather conditions and may adapt better through auxiliary tasks and fine-tuning.
  • The models have difficulty predicting very sharp photovoltaic-output changes caused by sudden cloud movement.The authors attribute this partly to insufficient visual differences between nearby input images and suggest shorter exposures.

8. Conclusion

The paper uses deep learning for photovoltaic nowcasting with past images and power outputs, emphasizing temporal dynamics, auxiliary tasks, and sky images. It reports success while noting that changing clouds still cause inaccurate predictions and motivates future integration with explicit cloud modeling and live adaptation.

  • Deep learning predicts future photovoltaic power from past photovoltaic outputs and sky images using an LSTM-based network.The approach outperforms simple baselines and more sophisticated multilayer-perceptron architectures.
  • The experiments identify temporal modeling, auxiliary learning tasks, and sky images as important aspects of the approach.These three components are presented as contributing aspects of the reported success.
  • Changing clouds remain difficult to model and can create inaccurate future photovoltaic-output predictions.The conclusion proposes combining end-to-end deep learning with explicit cloud-motion methods such as optical flow.
  • Future work includes deploying the method in a live solar-panel system and continuously learning from captured data.Active learning is suggested as a way to adapt to the properties of the solar panel used.
Loading 1810.06327v1…