Solar Intelligence

1 Introduction

• Solar energy planning and maintenance require domain knowledge, current research, available data, and forecasts to assess systems holistically.Installation planning also depends on long-term irradiance variability, seasonal trends, and unusual-pattern detection.Source: §1 · p. 1
• Fragmented dashboards, forecasting services, and language models make it time-consuming to gather research and data before making solar decisions.Traditional statistical and machine-learning models analyze historical patterns but do not support natural-language querying or evidence-grounded explanations.Source: §1 · p. 2
• Solar Intelligence brings solar-domain querying, analytics, and forecasting into one platform, reducing the need to move between separate tools.Its objectives include shared access to numerical and published evidence, cited natural-language answers, and forecasting through a dashboard and API.Source: §1 · p. 2

2 Data Sources

• The platform integrates 149 research papers and reports, daily NASA POWER records, and minute-resolution Biosphere 2 ground-sensor readings.The corpus spans 1994–2026; NASA POWER data covers 2012–2026, while the Biosphere 2 sensor record covers August 2012–July 2013.Source: §2 · p. 3, Table 1 · p. 3
• NASA POWER provides 5,225 daily records for Biosphere 2, covering irradiance, air temperature, relative humidity, wind speed, and precipitation.Its solar radiation comes from CERES SYN1deg, and its meteorological parameters from NASA’s MERRA-2 reanalysis.Source: §2 · p. 3
• Biosphere 2 ground sensors contribute 10.4 million minute-resolution readings of air temperature, solar irradiance, and wind speed.The readings cover August 2012 through July 2013.Source: §2 · p. 3

3 Methodology

• The methodology has five components: corpus construction, query routing and retrieval, ground-sensor cleaning, cross-source comparison, and forecasting.Each component is described in a separate methodology subsection.Source: §3 · p. 3

3.1 Corpus Construction and Processing

• The solar-energy corpus combines arXiv papers queried across 18 subdomains with institutional reports, and its PDFs yield 9,496 approximately 500-token chunks with 50-token overlap, indexed in parallel with BM25 and 384-dimensional BGE-small embeddings.BM25 tokenization uses lowercase normalization; stemming is omitted to protect domain-specific tokens.Source: §3.1 · pp. 3, 5

3.2 Retrieval through query routing

• A rule-based classifier routes numerical queries to DuckDB SQL, conceptual queries to hybrid retrieval, and combined queries through both paths.BM25 and dense-embedding results are fused with Reciprocal Rank Fusion; numerical outputs can bypass the language model and be constructed deterministically from SQL results.Source: §3.2 · p. 5

3.3 Biosphere 2 Data Cleaning

• Biosphere 2 measurements with implausible values are range-filtered to retain physical consistency.The authors identify sensor faults, communication errors, and unit-conversion artifacts as possible causes of implausible readings.Source: §3.3 · p. 5

3.4 NASA POWER and Biosphere 2 Data Comparison

• The comparison evaluates NASA POWER against cleaned Biosphere 2 readings over 312 overlapping days for irradiance, temperature, and wind speed.The reported metrics are each source’s mean, signed bias, and Mean Absolute Error.Source: §3.4 · pp. 5–6

3.5 Forecasting

• XGBoost produces daily irradiance, temperature, and wind-speed forecasts, trained only on NASA POWER data because Biosphere 2’s 11-month record is insufficient for daily-resolution forecasting requiring multi-year seasonal cycles.Source: §3.5 · p. 6

4 System Architecture and Implementation

• The system has four modular layers: data, analytics and knowledge retrieval, forecasting, and interface.The technology stack for these components is summarized in Table 4.Source: §4 · p. 6, Table 4 · p. 8

4.1 Data Layer

• The data layer stores NASA POWER and Biosphere 2 solar and meteorological data in DuckDB alongside unstructured scientific text.Structured variables include irradiance, temperature, humidity, wind speed, and precipitation.Source: §4.1 · p. 6

4.2 Analytics and Knowledge Retrieval Layer

• The retrieval layer routes structured questions to SQL, scientific questions to document retrieval, and hybrid questions to both.It combines BM25 and semantic embeddings with Reciprocal Rank Fusion, then grounds generated answers in structured outputs and retrieved evidence.Source: §4.2 · pp. 6–7

4.3 Forecasting Layer

• The forecasting layer uses XGBoost with historical data and engineered lag, rolling-average, and seasonal features.For multi-day forecasts, it iteratively feeds each prediction into the series before predicting the next day.Source: §4.3 · p. 7

4.4 Interface and API Layer

• FastAPI provides backend APIs, while Streamlit presents natural-language queries, data comparisons, forecasts, and evidence panels.The interface serves technical users and software integration.Source: §4.4 · p. 7

4.5 System Workflow

• The workflow classifies a natural-language query, selects structured data, scientific documents, or both, then presents the evidence-backed response in Streamlit.Users can also select a solar-energy variable for forecasting.Source: §4.5 · p. 8

4.6 Deployment Architecture

• Deployment uses separate FastAPI and Streamlit servers, with a daily NASA POWER data-loader job and an MCP-style manifest advertising nine callable tools.The API server has 13 REST endpoints, and the dashboard has five tabs.Source: §4.6 · p. 8

5.1 Query Types and QA Citation Results

• The query layer distinguishes semantic, structured, and hybrid questions; the hybrid example asks for a NASA–Biosphere 2 irradiance comparison with an explanation.The section presents outputs for semantic and structured queries.Source: §5.1 · p. 8

5.1.1 Semantic Query

• Semantic retrieval matched the 15%–22% commercial-panel efficiency range [1] but omitted the 0.4% penalty per degree above 25°C [1] and foregrounded a theoretical 69% ceiling [2].The paper notes that the theoretical bound was not distinguished from measured commercial efficiency, risking confusion for present-day efficiency queries.Source: §5.1.1 · p. 9

5.1.2 Structured Query

• Structured queries returned average GHI of 5.95 kWh m−2 day−1 for March 2025 and average Biosphere 2 wind speed of 1.258 m s−1 for March 2013 directly from SQL.The wind-speed result came from a DuckDB AVG() over the Biosphere 2 solar table.Source: §5.1.2 · pp. 10–11

5.2 NASA POWER and Biosphere 2 Comparison

• NASA follows seasonal patterns across all three variables, with strongest agreement for irradiance, a cold bias for temperature, and the weakest agreement for wind speed, which it overestimates.The comparison uses monthly averages from August 2012 to July 2013.Source: Figure 1 · p. 11, §5.2 · p. 11

5.3 Forecast Outputs

• Seven-day XGBoost roll-forward forecasts show a mid-week wind dip, near-flat irradiance, and temperature cooling followed by warming; evaluation against real values is still required.The reported forecast trajectories were judged to be in valid ranges, but the paper calls for detailed interpretation and evaluation before relying on them.Source: Figure 2 · p. 12, §5.3 · p. 13

• Forecasts are restricted to Biosphere 2, rely on NASA POWER data lagging by about three days, and provide point forecasts that limit reliable grid-operation use.The paper identifies the location restriction, data lag, and lack of uncertainty forecasts as limitations of the current system.Source: §6 · p. 13

• The authors present Solar Intelligence as a single system combining evidence-based retrieval over research and data with XGBoost forecasting, intended to bring fragmented resources together and save time.They identify a larger corpus, multi-site generalization, and improved uncertainty-aware forecasting as future work.Source: §7 · p. 13