晨光
暗夜
晨光
极光
Bilingual Paper Reading · 中英对照精读

SolarBoost:时变电网容量下的分布式光伏功率预测

准大一 · 电气工程 × 新能源 × AI+电力 —— 分布式光伏功率预测精读材料
原文:arXiv:2510.21129 2025年10月24日发布 arXiv 预印本(cs.LG) 分布式光伏 × 容量解耦 × 梯度提升树 附英文摘要朗读音频

一、论文档案

英文标题SolarBoost: Distributed Photovoltaic Power Forecasting Amid Time-varying Grid Capacity
中文标题SolarBoost:时变电网容量下的分布式光伏功率预测
作者耿林源, 杨林晓, 顾欣悦, 孙亮(达摩院 达摩院 团队)
发布时间2025年10月24日(v1)|分类:cs.LG(机器学习)
一句话概括把分布式光伏的总功率拆成「单位容量出力 × 动态容量」两件事分别建模——容量变了模型不用重训,已在中国的多个城市落地部署。
💡 为什么选这篇给你:① 分布式光伏(屋顶光伏)是中国新能源装机增长的主力,预测它的功率是电网消纳的刚需;② 思路极妙——「解耦」容量与单位出力,一个模型吃遍所有城市;③ 有理论证明、有真实部署数据(含弃光损失金额),故事完整、可复现(开源代码在 GitHub)。

二、核心术语表(先扫一遍再读正文)

英文术语中文大白话解释
distributed photovoltaic (DPV)分布式光伏装在屋顶、社区的小型光伏电站,数量多、分布散、数据杂。
centralized photovoltaic (CPV)集中式光伏并网的大型光伏电站,设备统一、特性均匀,好建模。
installed capacity装机容量某个区域所有光伏设备额定功率的总和(MW 级)。
time-varying grid capacity时变容量分布式光伏装机随时间不断增长,预测时「容量」本身在变。
unit output function单位容量出力函数「每 1 MW 容量在给定天气下能发多少电」的函数,与容量大小无关。
generalized capacity广义容量模型给每个网格估出的「容量」乘子,吸收设备差异、数据缺失等一切异质性。
decoupling解耦把「单位出力函数」和「容量」分开建模,互不干扰,各自动态适应。
spatial grid空间网格把预测区域划分成的小格子(注意:这里的 grid 不是电网)。
boosting trees提升树XGBoost、LightGBM 这类梯度提升树模型,中小型表格数据上的王者。
upper-bound approximation上界近似用一个数学上界去近似复杂的损失函数,绕开嵌套求和的算力瓶颈。
concept drift概念漂移数据分布随时间变化(比如装机容量增长),老模型逐渐失效的现象。
curtailment弃光(限功率)因电网消纳能力不足,调度强制光伏少发电——太阳白晒、钱白丢。
surrogate loss代理损失不直接优化原目标函数,而是优化一个计算上更可行的近似目标。
latent variable隐变量没有真实标签、由模型内部估计出来的变量(这里是广义容量)。

三、摘要中英对照(精读核心)

🎧 音频在文末,可先听一遍原文再读;每个英文句都配了逐句翻译。

摘要 Abstract

EN · 原文
This paper presents SolarBoost, a novel approach for forecasting power output in distributed photovoltaic (DPV) systems.
CN · 翻译
本文提出 SolarBoost——一个用于分布式光伏(DPV)系统功率输出预测的新方法。
EN · 原文
While existing centralized photovoltaic (CPV) methods are able to precisely model output dependencies due to uniformity, it is difficult to apply such techniques to DPV systems, as DPVs face challenges such as missing grid-level data, temporal shifts in installed capacity, geographic variability, and panel diversity.
CN · 翻译
现有的集中式光伏(CPV)方法因设备统一,能精确建模出力依赖关系;但把它们用到 DPV 上很困难,因为分布式光伏面临电网级数据缺失、装机容量时变、地理差异大、组件五花八门等挑战。
EN · 原文
SolarBoost overcomes these challenges by modeling aggregated power output as a composite of output from small grids, where each grid output is modeled using a unit output function multiplied by its capacity.
CN · 翻译
SolarBoost 把聚合总功率建模为若干小网格出力的合成:每个网格的出力 = 单位容量出力函数 × 该网格容量
EN · 原文
This approach decouples the homogeneous unit output function from dynamic capacity for accurate prediction.
CN · 翻译
这一方法把同质化的单位出力函数动态容量解耦开来,从而实现准确预测。
EN · 原文
Efficient algorithms over an upper-bound approximation are proposed to overcome computational bottlenecks in loss functions.
CN · 翻译
论文提出基于上界近似的高效算法,以攻克损失函数中的计算瓶颈
EN · 原文
We demonstrate the superiority of grid-level modeling via theoretical analysis and experiments.
CN · 翻译
我们通过理论分析实验证明了「网格级建模」的优越性。
EN · 原文
SolarBoost has been validated through deployment across various cities in China, significantly reducing potential losses and provides valuable insights for the operation of power grids.
CN · 翻译
SolarBoost 已在中国多个城市部署验证,显著减少了潜在损失,并为电网运行提供了有价值的见解。
EN · 原文
The code for this work is available at https://github.com/达摩院-DI-ML/SolarBoost.
CN · 翻译
本工作代码开源:github.com/达摩院-DI-ML/SolarBoost

关键词 Keywords:Distributed Photovoltaic Forecasting 分布式光伏预测 | Time-varying Capacity 时变容量 | Boosting Trees 提升树 | Capacity Decoupling 容量解耦

四、引言精选(为什么这个问题重要)

① 光伏为什么重要:一组数字说明一切

EN · 原文
Photovoltaic (PV) power converts sunlight directly into electricity using PV panel arrays, which is considered cost-effective, secure, and environmentally friendly. Consequently, PV generation is rapidly expanding, now accounting for 5.49% of global electricity in 2023, marking a 35.9% year-on-year increase in its share (Institute et al., 2024). There are two types of PV systems. Centralized PV (CPV): solar farms connected to the main power grids. Distributed PV (DPV): small installations on rooftops or in local communities. DPV systems are easy to install, require little maintenance and can utilize unused places such as rooftops or walls, making up 41% of global PV capacity (Agency, 2024) and helping make electricity available, clean and affordable for everyone.
CN · 翻译
光伏(PV)用光伏板阵列把阳光直接转化为电能,被认为经济、安全、环保。因此光伏发电快速扩张:2023 年已占全球电力的 5.49%,占比同比上升 35.9%。光伏分两类:集中式(CPV,并网大电站)与分布式(DPV,屋顶或社区小装置)。DPV 安装容易、维护少、能用屋顶和墙面等闲置空间,占全球光伏容量的 41%,让电更普及、更清洁、更便宜。

② 预测不准的代价:加州大停电与 4200 万美元损失

EN · 原文
Accurate forecasting of PV power output is essential for maintaining grid stability and maximizing the use of renewable energy. If solar power output is overestimated or underestimated, it can disrupt the balance between electricity supply and demand, sometimes leading to equipment overloads, forced curtailment, or even blackouts. For example, in August 2020, California experienced its largest rolling blackout since the 2001 energy crisis, due to miscalculations in balancing energy generation and demand (Macy, 2020). Another example is a publicly listed Chinese wind power company (power Corporation, 2023), which reported that in the first half of 2023, 6,440 million kilowatt-hours, or 9.1% of its generated power, were curtailed due to inadequate management of energy absorption. This is equivalent to 212,000 tonnes of standard coal waste, causing a direct economic loss over 42 million USD to this single company.
CN · 翻译
准确的 PV 功率预测对维持电网稳定、最大化利用可再生能源至关重要。高估或低估都会打破供需平衡,导致设备过载、被迫弃光甚至停电。例如 2020 年 8 月加州经历了自 2001 年能源危机以来最大规模的滚动停电;又例如一家中国上市风电公司报告:2023 年上半年因消纳管理不足被弃掉 6,440 百万千瓦时(占发电量 9.1%),相当于浪费 212,000 吨标准煤,给这一家公司造成超 4200 万美元的直接经济损失。

③ 分布式光伏的四大挑战:数据缺、容量变、地理散、设备杂

EN · 原文
Unfortunately, the majority of solar forecasting methods have been developed for CPV systems (Chu et al., 2024; Rahdan et al., 2024), which have relatively uniform characteristics. In contrast, forecasting for DPV systems faces additional challenges, including: (1) Limited and Noisy Data: Data from individual households or small-scale installations is often inconsistent, incomplete, or outdated, leading to difficulty in developing reliable models. (2) Temporal Scale Shifts: The capacity of DPV installations changes over time, introducing concept drift issues (Agrahari and Singh, 2022).
CN · 翻译
遗憾的是,大多数光伏预测方法都是为特性相对统一的 CPV 设计的。DPV 预测面临额外挑战:(1)数据有限且嘈杂——单个家庭/小规模装置的数据常常不一致、不完整或过时;(2)时间尺度漂移——DPV 装机容量随时间变化,带来概念漂移问题。
EN · 原文
(3) Geographical Diversity: DPV systems are widely scattered, experiencing different weather conditions. The approach that aggregates meteorological data to predict overall output as a single entity loses accuracy. Conversely, the method that predicts power data at consumer resolution is also less feasible due to limited and noisy data issues, as well as computational cost (Rahdan et al., 2024). (4) Equipment Diversity: Unlike CPV systems with standardized equipment, DPV systems feature a mix of panel types, orientations and maintenance statuses, making it hard for consistent modeling.
CN · 翻译
(3)地理多样性——DPV 散布各地、天气各异:把气象数据聚合成一个整体来预测会损失精度;而按用户级分辨率预测又因数据少而杂、算力成本高而不可行。(4)设备多样性——不像 CPV 设备标准化,DPV 是各种面板类型、朝向、维护状态的混合体,难以一致建模。

④ 核心思想:把「单位出力」和「容量」解耦

EN · 原文
To address these challenges of DPV power forecasting, we propose SolarBoost, a novel boosting-based approach for modeling DPV power output. SolarBoost predicts total power by decomposing the area into small spatial grids and inferring two quantities for each grid: (1) a generalized (dynamic) grid-level capacity and (2) a per-unit-capacity output function that links weather to output. By decoupling capacity from the unit output function, our model can adjust to capacity changes with a consistent output function, thus improving forecast accuracy.
CN · 翻译
为解决上述挑战,我们提出 SolarBoost:把区域分解为小空间网格,对每个网格估计两个量:(1)广义(动态)网格级容量(2)连接天气与出力的单位容量出力函数。通过把容量与单位出力函数解耦,模型能以一致的出力函数适应容量变化,从而提升预测精度。

⑤ 为什么选树模型:表格数据上的经验之谈

EN · 原文
Extensive business experience and prior research have shown that ensemble trees, particularly boosting frameworks like XGBoost and LightGBM, perform well with medium-sized tabular datasets and have outperformed deep learning methods in PV forecasting tasks (Hong et al., 2020; Rahimi et al., 2023). This motivates our exploration of innovations based on the gradient-boosted trees paradigm.
CN · 翻译
大量业务经验和已有研究表明:集成树(尤其是 XGBoost、LightGBM 等提升框架)在中小规模表格数据上表现优异,在光伏预测任务中甚至胜过深度学习方法。这促使我们在梯度提升树范式上进行创新。
💡 这是全文最有味道的一句“By decoupling capacity from the unit output function, our model can adjust to capacity changes with a consistent output function.”——「容量会变,但每 1 MW 的发电规律不变」。把不变的和变的分开建模,是这篇论文的第一性原理。

五、论文贡献(5 个要点)

EN · 原文
1. Capacity Modeling. We develop a framework to explicitly model dynamic, heterogeneous grid-level generalized-capacities, which serves as a multiplier for the output capturing all heterogeneity, including capacity changes (challenge 2), panel diversity (challenge 4) and missing grid-level information (challenge 1).
CN · 翻译
1. 容量建模。显式建模动态、异构的网格级广义容量——它作为出力的乘子,吸收一切异质性:容量变化、组件多样性和数据缺失。
EN · 原文
2. Consistent Unit Output Function. We introduce a consistent unit output function, ensuring stable weather-to-output modeling even as capacities evolve (challenge 2). Thus we can use broad data without sacrificing granularity (challenge 3).
CN · 翻译
2. 一致的单位出力函数。引入稳定一致的单位出力函数,即使容量不断演化,天气→出力的建模也保持稳定,从而能用大范围数据而不牺牲粒度。
EN · 原文
3. Efficient Optimization Framework on Boosting Trees. We present an efficient, scalable algorithm based on boosting trees with upper-bound approximations to tackle the computational difficulties in the loss function involving nested summations.
CN · 翻译
3. 提升树上的高效优化框架。提出基于提升树与上界近似的高效可扩展算法,攻克损失函数中嵌套求和带来的计算困难。
EN · 原文
4. Theoretical Proof for Accuracy. We provide theoretical analysis that substantiates the superiority of grid-based models over aggregated approaches. This analysis also demonstrates the adequacy of rough generalized-capacity estimations for accurate forecasts.
CN · 翻译
4. 精度的理论证明。给出理论分析,证明网格级模型优于聚合式方法;并说明「粗略估计广义容量就足够获得准确预测」。
EN · 原文
5. Deployment. SolarBoost has been deployed in various cities of China, showing its efficacy and adaptability in real-world DPV forecasting across different geographical contexts.
CN · 翻译
5. 落地部署。SolarBoost 已在中国多个城市部署,展示了其在不同地理环境下真实 DPV 预测中的有效性与适应性。

六、结论中英对照

EN · 原文
In this paper, we introduce SolarBoost, a novel method for DPV power forecasting. We investigate the effects of capacity fluctuations and the limitations associated with neglecting them. Additionally, we formally define the DPV forecasting problem within the context of time-varying grid capacities. We then propose the SolarBoost method, which operates iteratively by alternating updates between the unit-output predictor and the generalized capacity with the objective of minimizing a surrogate loss function. Although the generalized capacity variable does not require a ground truth label, it functions as an essential latent variable in our modeling, allowing the unit-output function to better capture the complex output mechanisms and meanwhile adapting to the change of total capacity. To the best of our knowledge, SolarBoost is the first approach to explicitly incorporate generalized capacity into the model. Experiments conducted on both toy and real-world datasets demonstrate that our method effectively forecasts generalized capacities and achieves low errors in both unit capacity outputs and aggregated power predictions. The explicit operational gains achieved through deployment are enormous, particularly on the reduction in power curtailment and increased revenue. These results demonstrate both the effectiveness of technical contributions and the real-world value in actual grid management and operations.
CN · 翻译
本文提出 DPV 功率预测新方法 SolarBoost。我们研究了容量波动的影响及忽视它的局限,并在时变容量背景下正式定义了 DPV 预测问题。SolarBoost 迭代地交替更新「单位出力预测器」与「广义容量」,以最小化代理损失为目标。广义容量虽然不需要真实标签,却是建模中至关重要的隐变量,让单位出力函数更好捕捉复杂出力机制、同时适应总容量变化。据我们所知,SolarBoost 是第一个显式把广义容量纳入模型的方法。在玩具数据集与真实数据集上的实验表明:它能有效预测广义容量,单位出力与聚合功率预测误差都很低。部署带来的运行收益巨大,尤其在减少弃光、增加收入方面。这些结果既证明了技术贡献的有效性,也体现了其在真实电网管理与运行中的价值。
EN · 原文
One limitation is that SolarBoost currently relies solely on DPV data. Incorporating CPV information and transferring their characteristics to the DPV context could achieve better accuracy. Future work will also focus on developing more effective optimization algorithms, modeling capacity with greater complexity, and extending SolarBoost's application to other renewable energy systems.
CN · 翻译
一个局限是:SolarBoost 目前只依赖 DPV 数据。若引入 CPV 信息并将其特性迁移到 DPV 场景,精度还能更高。未来工作还将聚焦:开发更有效的优化算法、以更高复杂度建模容量,以及把 SolarBoost 推广到其他可再生能源系统

七、编者解读:这篇论文到底讲了什么(大白话版)

  1. 问题:分布式光伏(屋顶光伏)越来越多,但它的数据烂(缺、杂、旧)、装机一直在涨、天气各地不同、组件型号五花八门。老办法(集中式光伏那套)直接套不上,预测不准就要多备火电、甚至弃光亏钱。
  2. 做法:SolarBoost 把一个大区域切成小网格,对每个网格只估两样东西——「这个格子里装了多少 MW(广义容量)」和「每 1 MW 在什么天气下发多少电(单位出力函数)」。总功率 = 各网格容量 × 单位出力之和。
  3. 妙处(解耦):容量爱怎么变就怎么变,单位出力函数始终保持一致——新装了光伏,只需更新容量,模型不用推倒重来。论文还用数学证明了「网格级建模」优于「整片聚合」。
  4. 工程细节:损失函数里有嵌套求和、算不动,作者用「上界近似」把计算变快;模型选梯度提升树(XGBoost/LightGBM 路线)而非深度学习,因为光伏表格数据上树模型更稳。
  5. 结果:已在中国多个城市部署,显著减少弃光损失(引言里提到弃电 9.1% 相当于 4200 万美元损失);代码开源可复现。
🎯 对保研的启示:这篇论文示范了「把难问题拆成简单问题」的工程思维——不变的规律(单位出力)与变化的量(容量)分开建模。复试时讲「我如何识别出问题中哪些部分稳定、哪些部分漂移,并分别建模」,比背模型名字高级得多。

八、给准大一的阅读路线图 & 延伸方向

📖 怎么读这篇论文(三遍法)

  1. 第一遍(10 分钟):只读摘要和术语表,回答三个问题——问题是什么?方法是什么?结果是什么?(答案:分布式光伏预测难;网格级容量×单位出力解耦;中国多城市部署见效。)
  2. 第二遍(20 分钟):读引言 + 结论,重点体会「四大挑战」和「解耦」这两个核心概念,再看结论坦诚的局限(只用 DPV 数据)。
  3. 第三遍(30 分钟):读方法文字部分(网格分解、广义容量、上界近似),跳过公式和引用编号,只看文字描述;遇到不懂的术语回查术语表。

🚀 这个方向你能延伸做什么

九、英文摘要朗读(练听力用)

先盲听一遍→再看对照稿→再听一遍。目标是听出每个数字(5.49%、35.9%、41%、42 million USD)和术语(distributed photovoltaic、generalized capacity、decouple、upper-bound approximation)。