Scholarlink

<p>We propose a forecasting and validation framework for log-changes in solar and wind power generation, with a focus on large forecast deviations relevant for power system balancing. Forecast accuracy is evaluated using standard point-error metrics, while extreme deviations are assessed through a Value-at-Risk (VaR)-based analysis of absolute forecast errors. In this setting, VaR is employed as a reliability-oriented validation measure: the unconditional frequency of threshold violations assesses calibration at extreme quantiles, while their temporal occurrence may reflect persistence in renewable generation shocks that can require operational actions such as curtailment or reserve activation. Using Terna’s 2023 data, we compare a physics-informed stochastic model with benchmark approaches including ARIMA-GARCH, SARIMA, Vasicek, Naive, LSTM, Random Forest and XGBoost models. Results show that the proposed model delivers the strongest point-forecast performance among the considered specifications and maintains comparable calibration at operational quantile levels. Clustering of threshold violations is observed across all models and is consistent with persistence in renewable generation regimes.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Risk-based validation of renewable energy forecasting: a test case for Italy

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Giacomo Ascione,
  • Michele Bufalo,
  • Giuseppe Orlando

摘要

We propose a forecasting and validation framework for log-changes in solar and wind power generation, with a focus on large forecast deviations relevant for power system balancing. Forecast accuracy is evaluated using standard point-error metrics, while extreme deviations are assessed through a Value-at-Risk (VaR)-based analysis of absolute forecast errors. In this setting, VaR is employed as a reliability-oriented validation measure: the unconditional frequency of threshold violations assesses calibration at extreme quantiles, while their temporal occurrence may reflect persistence in renewable generation shocks that can require operational actions such as curtailment or reserve activation. Using Terna’s 2023 data, we compare a physics-informed stochastic model with benchmark approaches including ARIMA-GARCH, SARIMA, Vasicek, Naive, LSTM, Random Forest and XGBoost models. Results show that the proposed model delivers the strongest point-forecast performance among the considered specifications and maintains comparable calibration at operational quantile levels. Clustering of threshold violations is observed across all models and is consistent with persistence in renewable generation regimes.