HouYi: An open-source large language model specially designed for renewable energy and carbon neutrality field

1. Introduction

• Fossil-fuel overuse is linked to environmental problems, while renewable energy is framed as important for addressing them and achieving carbon neutrality.The introduction says renewable energy development is accelerating with the carbon-neutrality goal.Source: §1 · p. 2
• The introduction motivates LLMs as tools for renewable-energy researchers to generate academic-paper paragraphs and improve writing efficiency.Source: §1 · p. 3
• The introduction identifies a lack of open-source renewable-energy LLMs and training datasets as restricting LLM applications in the field.It also notes that general LLMs may lack accurate domain-specific expert knowledge.Source: §1 · pp. 3–4
• HouYi showed better performance than general LLMs and excellent renewable-energy academic paper generation ability.The introduction reports these results without giving quantitative comparisons.Source: §1 · pp. 4–5
• REAP is presented as the first open-source renewable-energy dataset for LLM training, containing titles and abstracts of 1,168,970 Web of Science papers.The dataset is designated for non-commercial LLM research.Source: §1 · pp. 4–5
• HouYi is presented as the first renewable-energy LLM, developed by finetuning general LLMs, with a public web demo available to readers.The introduction describes it as an open-source LLM specially designed for the field.Source: §1 · pp. 4–5
• The paper introduces human-expert-based analytic hierarchy process (AHP) evaluation of LLMs and reports it as a first use of AHP for this purpose.Source: §1 · p. 5

• The Web of Science search used related keywords covering sustainable energy, renewable-energy technologies, energy storage, carbon neutrality, and related concepts.Source: §2 · p. 6
• 1,168,970 academic literatures were collected for REAP, using titles as LLM inputs and abstracts as outputs to create the first open-source dataset for non-commercial renewable-energy LLM research.Source: §2 · p. 6
• 711,126 journal papers are included, with the paper presenting their top 20 journal titles and article counts alongside figures on literature types and journal rankings.Source: §2 · p. 6, Figure 1 · p. 7, Figure 2 · p. 7

3.1. Large language model

• LLMs learn from large text collections and generate human-like responses; ChatGPT’s academic-writing assistance motivates an open-source renewable-energy model.The paper characterizes ChatGPT as a closed-source general LLM without sufficient renewable-energy focus.Source: §3.1 · p. 8, Figure 3 · p. 8

3.2. Transformer

• Transformer [34] uses stacked encoder layers with self-attention and feed-forward sublayers, while decoder layers add attention over encoder outputs.Residual connections and layer normalization are used in the encoder’s add-and-norm component.Source: §3.2 · pp. 9–10

3.3 General Language Model (GLM)

• GLM [35] supports understanding and conditional or unconditional generation; its pretraining objective fills missing token spans in corrupted text.The paper notes that multi-task pretraining enables GLM to perform well across various tasks.Source: §3.3 · p. 11

3.4. Finetune transfer learning

• Because general GLM lacks sufficient renewable-energy focus, the authors fine-tuned ChatGLM-6B with REAP to develop HouYi for the field.The paper describes HouYi as the first LLM for renewable energy.Source: §3.4 · pp. 12–13
• The paper fine-tuned ChatGLM-6B on REAP title-and-abstract pairs, with titles as inputs and abstracts as outputs, for 30 epochs at learning rate 0.01 on Ubuntu 20.04 using eight RTX 3090 graphics cards.Source: §3.4 · p. 13
• The case study models a 2,000 MW plant with 30 turbines rated at 100 MW each, where wind speed varies from 30 to 50 m/s.The stated temperature ranges are 10–40°C in summer and 10–30°C in winter.Source: §3.4 · p. 15
• The company’s output-power model weights input power at 0.8 and wind speed and power factor at 0.1 each, and was run 100 times.Output power and power-factor uncertainties are reported as 0.5% per run, with 1% uncertainty also stated for the first and second runs.Source: §3.4 · p. 15
• The method is presented as applicable in other countries for estimating uncertainty in the wind-energy conversion process.The case study is described as illustrating the probabilistic power-curve modelling method.Source: §3.4 · p. 15
• The method uses probabilistic power-curve modelling to estimate uncertainty in wind energy conversion and support more reliable turbine-performance predictions.The passages also describe using the method to optimize turbine output based on input power and weather conditions.Source: §3.4 · pp. 15, 17
• The passages identify randomness and uncertainty in wind speed as a difficulty for power-curve modelling and a limitation of deterministic methods.The stated aim is to explore probability-based modelling and evaluate uncertainty during modelling.Source: §3.4 · p. 18
• The wind power-curve workflow collects actual turbine output data, fits a Weibull distribution to wind speed and power generation using maximum likelihood estimation, and evaluates model performance.Source: §3.4 · p. 18
• The probabilistic wind power-curve model was reported to have good prediction accuracy and stability, and the study evaluated modeling uncertainty.Source: §3.4 · p. 18
• For the photovoltaic comparison, researchers used a real-world grid-connected PV case and a NOAA ANN forecasting model trained on historical PV-production and weather data.Source: §3.4 · pp. 20–21
• CNNs improved photovoltaic forecasting accuracy, SVMs did not provide significant improvement, and combining CNNs with SVMs gave the best performance.Source: §3.4 · p. 20
• The comparison considers deep-learning methods such as CNNs, SVMs, and traditional regression methods for improving power-forecasting accuracy.The passage identifies linear regression as an example of a traditional method.Source: §3.4 · p. 21
• The study evaluates the power-forecasting model using leave-one-out cross-validation and compares its performance with traditional methods.The described procedure trains and tests the model on separate datasets before evaluating test-set performance.Source: §3.4 · p. 21
• CNNs significantly improve one-day-ahead photovoltaic forecasting accuracy, SVMs also perform well, and their combination provides the best performance.The reported findings describe CNNs as improving accuracy and the combined CNN–SVM approach as best-performing.Source: §3.4 · p. 21

4.2 ChatGPT Assistant Evaluation

• ChatGPT rated each model on helpfulness, relevance, accuracy, and detail using a 0–100 overall score, with five evaluations averaged for each model.ChatGPT was used instead of GPT-4 because of GPT-4’s difficulty and expense.Source: §4.2 · pp. 25–26
• For both wind- and solar-energy paper generation, HouYi is comparable to ChatGPT, slightly outperforms Claude, ERNIE Bot, and SparkDesk, and significantly outperforms LLaMA-13B.Tables 5 and 6 present the ChatGPT evaluations for wind and solar generation, respectively.Source: Table 5 · p. 26, Table 6 · p. 26, §4.2 · p. 27

4.3 Human experts-based analytical hierarchy process for performance evaluation

• Five renewable-energy experts assessed wind- and solar-paper generation on helpfulness, relevance, accuracy, detail, academic standard, and experimental design.The results are reported separately for the two energy domains.Source: §4.3 · p. 27
• AHP weighted the six metrics, and its judgment matrix passed the consistency test with CR = 0.0831, below the 0.1 criterion.The maximum eigenvalue was 6.5232.Source: §4.3 · pp. 29–30
• Human-expert AHP scores place HouYi above Claude, ChatGPT, ERNIE Bot, and SparkDesk, and significantly above LLaMA-13B.The comprehensive scores are based on six weighted evaluation metrics.Source: §4.3 · p. 30, Table 20 · p. 30

Conclusion

• 1,168,970 Web of Science literatures make up the first open-source REAP dataset for non-commercial renewable-energy LLM research, pairing titles with abstracts.Source: §5 · p. 31
• Based on REAP, HouYi is developed by fine-tuning general LLMs as the first LLM specially designed for renewable energy.Source: §5 · p. 31
• The evaluation combines ChatGPT’s automatic assessment with human-expert AHP, which the authors identify as its first use for LLM evaluation.Source: §5 · p. 31
• HouYi achieves the highest scores in wind and solar paper generation, matches ChatGPT comparably, slightly outperforms Claude, ERNIE Bot, and SparkDesk, and significantly outperforms LLaMA-13B.Source: §5 · p. 31

Appendix 1 ChatGPT’s detailed evaluation for wind energy text generation task

• Assistant 4 gives a detailed explanation and modeling steps but is judged less relevant and detailed than earlier responses.Its response discusses probabilistic power-curve modeling, but some sections need further elaboration.Source: Appendix 1 ChatGPT’s detailed… · p. 33
• Assistant 5 provides a brief overview, whereas Assistant 6 gives only a title or concise basic introduction and lacks detail.Assistant 5’s response is described as less detailed and accurate than previous assistants; Assistant 6 is not considered highly valuable because it lacks explanation.Source: Appendix 1 ChatGPT’s detailed… · pp. 33–34
• The comparative evaluation considers helpfulness, relevance, accuracy, and response detail, and says response order does not affect the assessment.Source: Appendix 1 ChatGPT’s detailed… · p. 34
• Assistant 1 provides a well-structured, detailed response covering necessary sections; Assistant 2 offers a comprehensive outline covering necessary sections, while Assistant 3 presents a well-organized response with sufficient detail for a topic overview.Source: Appendix 1 ChatGPT’s detailed… · p. 34
• 95, 92, 90, 88, and 85 were assigned to Assistants 1, 3, 2, 4, and 5, respectively.The passage also lists 80 as the sixth score, but the supplied passages do not explicitly identify which assistant received it.Source: Appendix 1 ChatGPT’s detailed… · p. 35
• Assistants 2 and 4 gave accurate, relevant responses, but their coverage lacked some specific details found in other responses.Assistant 2 outlined paper sections and content, while Assistant 4 provided a concise overview of wind-turbine power-curve challenges and probabilistic methods.Source: Appendix 1 ChatGPT’s detailed… · p. 36
• Assistant 6 received a lower score because its brief response lacked structure, specific information, and sufficient detail.It mentioned uncertainty estimation and wind energy conversion but did not outline the paper’s main sections.Source: Appendix 1 ChatGPT’s detailed… · pp. 36–38
• Assistant 5 received a score of 92, compared with Assistant 6’s 80; Assistant 5’s response was described as well-structured, highly relevant, accurate, and including a case study.Source: Appendix 1 ChatGPT’s detailed… · p. 38
• In summary, all six AI assistants provide relevant information about uncertainty estimation for wind energy conversion through probabilistic wind turbine power curve modelling; Assistants 1, 3, and 5 stand out with high scores.Some responses offer more comprehensive and detailed explanations than others.Source: Appendix 1 ChatGPT’s detailed… · p. 38
• One response covered solar power forecasting background, status, machine-learning methods, applications, approach, and results, but lacked some details despite being helpful and relevant.It discussed how machine-learning algorithms can improve prediction accuracy and offered insight into method performance under different conditions.Source: Appendix 1 ChatGPT’s detailed… · p. 43
• Assistant 5 offered a comprehensive, relevant, accurate, and helpful outline spanning forecasting methods, evaluation, industry implications, limitations, and future research.Its proposed sections included abstract, introduction, literature review, data collection and preprocessing, results and analysis, and discussion.Source: Appendix 1 ChatGPT’s detailed… · p. 43
• Assistant 6 was unrelated to the question, expressed unfamiliarity, and linked a paper on controlled ensemble prediction without providing a comprehensive evaluation, earning lower scores.Its helpfulness, relevance, accuracy, and level of detail were rated lower.Source: Appendix 1 ChatGPT’s detailed… · p. 43
• Assistant 3 was rated most comprehensive and detailed, followed by Assistants 1, 5, and 2; Assistant 4 provided a good foundation, while Assistant 6 was irrelevant.Source: Appendix 1 ChatGPT’s detailed… · p. 43