Source-linked AI summary
Numerical Weather Prediction (NWP) and hybrid ARMA/ANN model to predict global radiation
Cyril Voyant, Marc Muselli, Christophe Paoli, Marie Laure Nivet
TL;DR
The paper develops a technique for forecasting hourly global radiation using ALADIN weather forecasts and a hybrid ANN/ARMA approach. Forecasting results can be improved by selecting ANN or ARMA models according to performance, while confidence intervals assess prediction reliability.
Problem
Hourly global radiation forecasting is addressed using meteorological forecasts from the ALADIN numerical weather prediction model.
Method
The approach optimizes a Multi-Layer Perceptron with ALADIN and stationary endogenous data, then combines it with an Auto-Regressive and Moving Average model.
Results
Forecasting results can be significantly improved by selecting ANN or ARMA models according to their performances.
Takeaways & Limitations
The study also computes confidence intervals to consider the reliability of the predictions.
Takeaways & Limitations
Out-of-sample or prediction performance will be poor when the process is modeled on a finite time interval.
Abstract
from arXiv · showhide
We propose in this paper an original technique to predict global radiation using a hybrid ARMA/ANN model and data issued from a numerical weather prediction model (ALADIN). We particularly look at the Multi-Layer Perceptron. After optimizing our architecture with ALADIN and endogenous data previously made stationary and using an innovative pre-input layer selection method, we combined it to an ARMA model from a rule based on the analysis of hourly data series. This model has been used to forecast the hourly global radiation for five places in Mediterranean area. Our technique outperforms classical models for all the places. The nRMSE for our hybrid model ANN/ARMA is 14.9% compared to 26.2% for the naïve persistence predictor. Note that in the stand alone ANN case the nRMSE is 18.4%. Finally, in order to discuss the reliability of the forecaster outputs, a complementary study concerning the confidence interval of each prediction is proposed
1. Introduction
The paper addresses hourly global-radiation forecasting for solar-energy applications by combining ALADIN weather forecasts with optimized ANN and ARMA models. It proposes a hybrid approach intended to improve predictions and report their confidence to end users.
- Motivation: Solar-radiation forecasting supports solar-project siting, electricity, heating and cooling applications, and grid stability.The motivation is especially relevant to solar-energy management in the Mediterranean area.
- Prior work: Earlier studies found optimized ANN and ARMA models produced similar global-radiation prediction performance.The earlier ANN used stationary endogenous inputs and exogenous meteorological data.
- Contribution: The proposed technique uses ALADIN weather forecasts as new inputs instead of previously used exogenous data.The paper assumes numerical weather prediction data can provide useful spatially distributed information for regional studies.
- Contribution: The method optimizes an MLP with stationary endogenous data and ALADIN forecasts, then combines it with ARMA using rules derived from hourly series.A pre-input layer selection method is used during architecture optimization.
- Evaluation: The hybrid model is evaluated at five Mediterranean locations and is reported to outperform classical models at all places.The study also computes confidence intervals to provide reliability information for forecast users.
2. Radiation time series and Numerical Weather Prediction
The study combines hourly radiation measurements from five Mediterranean coastal stations with meteorological forecasts from ALADIN-France. It uses selected forecast variables, nearby grid points, and a daily 06AM analysis with a 12-hour forecast horizon, while acknowledging scope and model-cycle limitations.
- Radiation data: The five stations are Ajaccio, Bastia, Montpellier, Marseille, and Nice, all located in the Mediterranean coastal area.The stations were maintained by Météo-France and equipped with pyranometers.
- Radiation data: Hourly global-horizontal radiation was measured from October 2002 through December 2008, using four years for model setup and two years for testing.Only observations from 8:00AM to 04:00PM true solar time were considered.
- Forecast data: ALADIN-France is a limited-area numerical weather prediction model with approximately 9.5 km horizontal resolution, 60 vertical levels, and a 54-hour forecast range.The model provides more than twenty outputs at one-hour temporal resolution, but not global radiation directly.
- Forecast data: The selected ALADIN predictors were pressure, nebulosity, rain precipitation, and temperature because they appeared directly linked to solar radiation.The nearest computational-grid points to the five stations were used.
- Forecast data: A 06AM analysis with a 12-hour forecast horizon was selected to cover the central prediction hours using one ALADIN extraction per day.The authors note that periodic changes in the ALADIN model may affect ANN training capability.
3. Forecasting models
The forecasting framework requires stationarity because ARMA and ANN models are treated as stationary-process modeling tools. The paper removes yearly and daily radiation seasonality, validates the transformation, and combines ARMA with ANN in the final prediction model.
- Hybrid model: The final forecasting model combines ARMA and ANN components after the radiation series has been transformed for stationary modeling.The proposed methodology is designed for a series with deterministic seasonal structure.
- Time-series prerequisites: Stationarity means that relevant statistical properties remain stable over time; weak stationarity requires time-invariant first and second moments.Strict stationarity additionally requires the full process distribution to be invariant under time shifts.
- Time-series prerequisites: The paper argues that stationarity is a modeling condition for both ARMA and neural-network forecasting.An ANN can mimic a non-stationary process over a finite interval, but its out-of-sample performance may be poor.
- Stationarity methodology: Global radiation is transformed through a clear-sky model, ratio-to-trend operations, moving-average seasonal correction, annual averaging, and construction of a new stationary series.The procedure targets both yearly and daily periodicities in the radiation series.
- Stationarity methodology: The stationarity workflow uses the Solis clear-sky model and periodic coefficients derived from moving-average corrections.The resulting methodology is summarized in Figure 2.
- Validation: A variation coefficient enables cross-comparison but cannot alone distinguish stationary from non-stationary processes.The study therefore also uses a Fisher variance-analysis test for daily and yearly seasonality.
b. ARMA
The ARMA/ANN framework uses stationary radiation data, optimized ARMA baselines, and an MLP whose inputs are selected through regression-based statistical testing. The hybrid design combines these forecasting components and reduces unnecessary ANN inputs.
- ARMA: ARMA models represent a stochastic process by coupling autoregressive and moving-average components, parameterized by p and q.The paper fits ARMA models after stationarizing and standardizing the global-radiation series.
- Hybrid forecasting: The forecasting pipeline first makes global radiation stationary and centered/reduced, then applies the hybrid ARMA/ANN methodology.Figure 4 depicts this hybrid forecasting process and identifies L as the lag or backshift operator.
- ARMA: After optimization, ARMA(1,0) was used for Ajaccio, Bastia, Montpellier, and Nice, while ARMA(2,0) was used for Marseille.The ARMA(2,0) prediction is expressed by Equation 16.
- ANN: The MLP predicts a future time-series value from a fixed number of past values and can incorporate endogenous radiation and exogenous ALADIN inputs.The implementation uses one hidden layer with hyperbolic-tangent hidden activation and linear output activation.
- Input selection: The pre-input selection method applies multiple linear regression and Student T-tests to choose useful variables from up to 10 endogenous lags and 2 lags per ALADIN output.The candidate pool contains 18 pre-input nodes: 10 endogenous and 8 exogenous.
- Input selection: Variables whose regression weights are statistically indistinguishable from zero are excluded from the neural-network input layer.This rule reduces the dimension of the input layer while following a parsimony principle.
4. Experiments, results and discussion
Experiments across five Mediterranean cities evaluate stationarity preprocessing, MLPs using endogenous and ALADIN data, and a hybrid ARMA/ANN forecast. The hybrid approach reduces average prediction error, while performance varies by season and location.
- Experiences about series makes Stationary: Stationarity experiments compare untreated radiation series with CSI and CSI* preprocessing using variation coefficients and Fisher tests.CSI* is reported as the most efficient preprocessing method across the studied cases.
- Experiences about series makes Stationary: CSI* produces the most non-seasonal series, while CSI retains stronger seasonal variation in the variance.The Bastia case illustrates the preprocessing effects on a difficult, noisy location.
- Results with MLP and ALADIN forecast data: Method V, an optimized MLP using ALADIN forecasts with CSI* preprocessing, is the most relevant method by annual error.Seasonally, ARMA performs better in four specified city-season cases, while other comparisons favor preprocessing-enhanced ANN variants.
- Results with MLP and ALADIN forecast data: 13.0% and 13.2% errors are reported when CSI* preprocessing and ALADIN data are added separately to ANN, while their combination further improves prediction.The paper attributes the most favorable prediction conditions to summer's abundant sunlight and low cloud occurrence.
- Results with the hybrid methodology: The hybrid predictor uses ARMA for one-third and ANN for two-thirds of hourly simulations, benefiting all examined seasons and cities.In summer, Marseille reaches an error below 10%.
- Results with the hybrid methodology: 11.3% is the average five-city nRMSE decrease against naïve persistence, whose average nRMSE is 26.2%.Measured and simulated radiation show strong correlation overall, but Bastia cloudy periods are not anticipated reliably.
- Results with the hybrid methodology: The method's reported result is limited to the particular ANN and ARMA models considered, and Bastia requires further improvement for nebulosity anticipation.The approach may also be complex and costly to implement.
- Results with the hybrid methodology: Confidence intervals complement the forecasts by indicating atypical or weakly supported predictions.The hybrid curve is visually most interesting, although its confidence interval remains only acceptable.
7. Conclusion
The paper combines an optimized MLP using ALADIN and stationary endogenous data with an ARMA model to forecast hourly global radiation at five Mediterranean locations. The hybrid approach reduces prediction error relative to standalone and persistence methods, while confidence intervals address forecast reliability.
- Method: The proposed method combines an optimized MLP using ALADIN forecasts and stationary endogenous data with an ARMA model selected from hourly-series analysis.An innovative pre-input layer selection method is also used.
- Method: The combined model successfully forecasts hourly global horizontal radiation for five Mediterranean locations.The study also establishes a stationarity process for global radiation time series before forecasting future insolation values.
- Results: 3.5% is the average nRMSE gain attributed to coupling ANN and ARMA predictors.Using ALADIN forecasting data as MLP input provides an additional average nRMSE gain of 0.7%.
- Implications: The study proposes confidence intervals to estimate the reliability of each prediction.The conclusion states that the method is compatible with renewable-energy deployment needs, particularly where interconnection is limited.
- Limitations: The model’s generalization remains to be demonstrated, including verification that superior performance is not caused by data mining or data snooping.The authors identify this as an important direction for establishing practical validity.
- Applications: The one-hour forecasting horizon matches the starting delay of conventional generators such as diesel engines and gas turbines.The paper identifies this horizon as important for electrical-grid managers handling renewable generation.
- Future work: Future work includes 24-hour-ahead solar-irradiance forecasting, testing on a real PV module, and simplifying implementation costs and complexity.The authors also propose connecting several MLPs to reduce the time step.
List of captions
The captions identify normalized CSI measures, compare ARMA with ANN+ALADIN+PC+CSI, and describe plotted measurement and forecast periods.
- CSI* and CSI are normalized in the referenced caption.
- One caption compares the ARMA methodology with ANN+ALADIN+PC+CSI over a 250-hour period.
- The referenced plotted periods span half-February to half-March 2008 and the first hours of January 2007.A caption states that the line corresponds to the measurement.