An accurate forecast of electric demand is essential for the optimal design of a generation system. For district installations, the projected lifespan may extend one or two decades. The reliance on a single-year forecast, combined with a fixed load growth rate, is the current industry standard, but does not support a multi-decade investment. Existing work on long-term forecasting focuses on annual growth rate and/or uses time resolution that is coarser than hourly. To address the gap, we propose multiple statistical forecast models, verified over as long as an 11-year horizon. Combining demand data, weather data, and occupancy trends results in a hybrid statistical model, i.e., generalized additive model (GAM) with a seasonal autoregressive integrated moving average (SARIMA) of the GAM residuals, a multiple linear regression (MLR) model, and a GAM with ARIMA errors model. We evaluate accuracy based on: (i) annual growth rates of monthly peak loads; (ii) annual growth rates of overall energy consumption; (iii) preservation of daily, weekly, and month-to-month trends that occur within each year, known as the 'seasonality' of the data; and, (iv) realistic representation of demand for a full range of weather and occupancy conditions. For example, the models yield an 11-year forecast from a one-year training data set with a normalized root mean square error of 9.091%, a six-year forecast from a one-year training data set with a normalized root mean square error of 8.949%, and a one-year forecast from a 1.2-year training data set with a normalized root mean square error of 6.765%.

翻译：电力需求的准确预测对于发电系统的最优设计至关重要。对于区域能源设施，其预期寿命可能长达一二十年。当前行业标准依赖于单一年度的预测并结合固定的负荷增长率，但这无法支撑长达数十年的投资决策。现有的长期预测研究主要关注年度增长率，和/或使用的时间分辨率粗于小时级别。为弥补这一不足，我们提出了多种统计预测模型，并在长达11年的预测范围内进行了验证。结合需求数据、天气数据和占用率趋势，我们构建了一个混合统计模型，即包含广义加性模型（GAM）及其残差的季节性自回归积分滑动平均模型（SARIMA）、多元线性回归（MLR）模型，以及带有ARIMA误差的GAM模型。我们基于以下方面评估预测精度：（i）月度峰值负荷的年度增长率；（ii）总能耗的年度增长率；（iii）对每年内发生的日、周及月际趋势（即数据的“季节性”）的保持能力；以及（iv）对全范围天气和占用条件下需求的真实反映。例如，模型使用一年的训练数据集进行11年预测，其归一化均方根误差为9.091%；使用一年的训练数据集进行6年预测，归一化均方根误差为8.949%；使用1.2年的训练数据集进行1年预测，归一化均方根误差为6.765%。