Solar power forecasts used by grid operators predict tomorrow's weather well and then assume the panels are clean. That assumption breaks in a predictable way. Dust builds up slowly and then washes off in a single rainstorm. Snow can erase a plant's output overnight and return it when the snow melts. Wildfire smoke dims the sky and also leaves particles on the glass. Each of these produces a forecast error that grows slowly and then flips sign suddenly, and each flip is tied to weather that is itself forecastable. Today's forecasters either fold these errors into "noise" or absorb them with a slowly adapting correction factor. The factor lags exactly at the moments that matter: the morning after rain, the day after a snowfall, the week of a smoke event.
We propose a camera-free forecaster that tracks a hidden "glass transparency" state for every plant in a fleet using only power output and weather data. The state is modeled as a saw-tooth process. It decays at a site-specific rate, drops abruptly with snow or heavy deposition, and resets with rain, melt, or cleaning. The model then pushes this state forward using the weather forecast, so a predicted storm produces a predicted cleaning. Explained to a colleague in two sentences: the forecaster learns how dirty each panel is from its own power data, and learns what makes it clean again. Because rain and snow are in the weather forecast, it can anticipate the jumps in output that current forecasters only notice a day late.
Core Claim
Separating a plant's output into three parts makes soiling a forecastable quantity rather than a residual to be absorbed. The parts are weather-driven output per unit of capacity, slowly changing installed capacity, and a surface-transmission state with explicit accumulate-and-reset dynamics driven by weather covariates. Two measurable consequences follow. First, systematic overprediction during soiled periods and systematic underprediction just after cleaning should both shrink, especially at day-ahead horizons where a reset can be anticipated from forecast precipitation. Second, pooling reset events across nearby plants should let one plant's observed cleaning inform its neighbors' state estimates, because rain is regional. This should help most in the dry, rain-sparse climates where single-site self-cleaning estimators fail.
The closest existing work connects soiling and forecasting in a very different way. Ka et al.'s multimodal forecaster segments soiling in close-range panel photographs and fuses it with sky-camera cloud classes and weather data. It raises single-step R² from 50.54% with weather alone to 81.79% with soiling features and 85.70% with all three streams. The authors state that such multimodal data is expensive to collect. Their soiling is also artificially placed (sand, pollen, plastic bags) and sampled weekly, so no accumulation or cleaning dynamics exist to forecast. We keep their motivating finding, that soiling state explains a large share of the forecast error, and discard the cameras. We replace them with a dynamical latent state that can be propagated forward over multi-hour and day-ahead horizons.
The Model
For site $s$ at time $t$, we write measured power as $P_{s,t} = c_s(t)\,\tau_{s,t}\,f(w_{s,t}) + \epsilon_{s,t}$. Here $f$ maps local weather $w$ (irradiance or its forecast, temperature, sun geometry) to output per unit of clean capacity, $c_s(t)$ is effective installed capacity, $\tau_{s,t}\in[0,1]$ is the fraction of light the glass lets through relative to clean, and $\epsilon$ is observation noise. The transmission state evolves as
$$\tau_{s,t+1} = \begin{cases} \tau_{s,t} - r_s(x_{s,t})\,\Delta t & \text{no event} \\ \tau_{s,t} + \rho(x_{s,t})\,(1-\tau_{s,t}) & \text{cleaning event} \\ \tau_{s,t}\,(1-\delta(x_{s,t})) & \text{covering event} \end{cases}$$
Here $x$ are driver covariates: precipitation, snowfall and temperature, wind, and airborne particulate concentration where available. $r_s$ is a site-specific deposition rate, $\rho$ is the fraction of soiling removed by a cleaning event, and $\delta$ is the fractional loss from snow or heavy deposition. Event probabilities are learned functions of $x$ with a spatially shared component. Capacity $c_s$ is restricted to rare, persistent upward steps plus slow monotone aging.
This restriction is what separates capacity from soiling: capacity never saw-tooths, and soiling resets coincide with weather drivers. It resolves a tension the file points to. SolarBoost tracks a latent "generalized capacity" with a Kalman filter to handle unmetered rooftop additions in distributed fleets. Its dynamics assume growth or gradual drift, so a soiling-and-rain cycle would be misread as capacity loss and recovery, and would corrupt the learned output function. We reuse SolarBoost's alternating structure, fitting $f$ by gradient-boosted trees given the latent states and then filtering the states given $f$. We change the state model from a random walk to a switching process. Filtering uses a particle filter over event indicators, with the continuous state handled in closed form within each particle.
The design borrows from two soiling-estimation lines. The data-driven soiling detection work estimates the soiling ratio from SCADA data (the plant's operational sensor logs) alone. It anchors clean baselines at rain-induced efficiency jumps (its forward and backward checking estimators) and reaches RMSE 0.005 against ground truth at Eugene, Oregon, versus 0.019 for the industry-standard Stochastic Rate and Recovery method. That result shows camera-free soiling estimation is feasible. The method is retrospective, site-by-site, and depends on rain, and we make it predictive and fleet-pooled. van Noord et al. infer tilt and azimuth from production data and build a snow-free reference across 258 Swedish sites (Pearson r = 0.96). They find annual snow losses up to 20% and show that the physical Marion snow model correlates at only 0.73. We treat snow as the covering branch of the same state, driven by forecast snowfall and temperature. Their work tells us production-data inference is reliable enough to anchor the state, and that a fixed physical snow model is not good enough to replace a learned one.
Research Program
The work has three parts. The first is the estimator: the switching state-space model above, fit jointly across a fleet. Spatial sharing enters only through the event model. A cleaning event is more likely at site $s$ if neighbors within a distance-weighted kernel show simultaneous jumps. We use kernel weighting in the style of the Haversine-distance graph of Karimi et al.'s spatiotemporal graph network, which showed that upstream plants give early warning of cloud fronts. Here the shared signal is a slow deposition-and-washing wave rather than a fast cloud. The second part is the forecaster. At each issue time we propagate the particle set forward under forecast drivers and combine it with any weather-to-power model. This yields a predictive distribution whose width reflects uncertainty about whether tomorrow's rain will clean the panels. The third part is analysis. We check whether the learned deposition rates and reset efficiencies are physically sensible, and quantify how much forecast error each regime (dust, snow, smoke) contributes.
The state model is deliberately attachable to any base forecaster. Its claim is about the multiplicative surface term, not about beating cloud-prediction models at cloud prediction. A satellite-plus-foundation-model system such as Baguan-solar or the st-GNN could supply $f$ unchanged.
Baselines
The naive reference is a forecaster that assumes clean panels. It uses the same gradient-boosted weather-to-power model, trained on all history, with no state term. This corresponds to the "clean panel invariant" practice. It shows how large the soiling-attributable bias is in each dataset and bounds what any correction can gain.
The strongest existing competitor on this exact question is SolarBoost's decoupled output-times-capacity model with a Kalman-filtered latent factor. It already has a multiplicative latent term that can absorb soiling if its process noise is tuned. SolarBoost reports 14% to 45% error reductions over CNN-LSTM and SARIMAX baselines on its eastern China dataset, and converges in about 46 seconds versus 30 minutes for the neural baselines. We will run it with the same gradient-boosted $f$, the same features, and the same data splits. Its Kalman process and observation noise will be tuned on the same validation data with the same search budget as our model's hyperparameters, so any gain comes from the switching dynamics and driver-conditioned events rather than extra tuning. Beating it means lower day-ahead bias and pinball loss (a standard scoring rule for quantile forecasts) on the days following cleaning or covering events, without being worse on event-free days.
For pure soiling-state estimation, the competitor is the backward checking estimator from the data-driven soiling detection work, with its reported RMSE of 0.005 at Eugene, Oregon. It sees the whole series, including the future, which our online filter does not. We therefore compare on two terms. Matching its RMSE within a small margin while running online counts as success. Beating it at dry-climate sites with few rain resets counts as a clear win.
For cloud-driven accuracy at intraday horizons, the reference is the st-GNN at a reported 120-minute MAPE of 11.01% on its 316-plant US fleet, versus 18.98% for a one-dimensional temporal convolution. We expect our state term to add a modest gain on top of it rather than replace it. We will report st-GNN alone and st-GNN multiplied by our propagated transmission state.
Two ablations isolate the key ingredient. The first replaces the driver-conditioned event model with event hazards that do not depend on weather. The model can still detect resets after the fact but cannot anticipate them. The second removes spatial pooling of events. If the first ablation matches the full model, anticipation adds nothing and the claim fails. If the second matches it, fleet pooling is unnecessary.
Why not just…?
Why not just use SolarBoost's Kalman filter with larger process noise? A random-walk multiplicative factor with enough noise will track soiling. This is the most serious threat to the proposal. Its limitation is structural. A random walk has no idea a reset is coming, so after an overnight rainstorm it overpredicts the dirty level for one or more days until the filter catches up. Raising the noise to make it catch up faster makes it chase every cloud-induced residual. Our model predicts the reset from forecast precipitation before it is observed. The decisive experiment is the post-event-day bias comparison described below. If a tuned Kalman filter comes within noise of our model there, we will report that the simpler model suffices and that anticipation is not worth its complexity.
Why not just add "days since last rain," snow depth, and particulate concentration as features to a gradient-boosted forecaster? A tree model can learn that power is higher after rain. This is a real alternative and could capture much of the day-ahead effect. What it cannot represent is site-specific accumulated state. Two sites with the same days-since-rain can differ in dirtiness because of different deposition rates, partial cleanings, or manual washing that appears in no feature. In arid climates, Anand notes that models calibrated elsewhere fail badly under heavy dust loading. There, days-since-rain grows without bound and conveys little. We will run this feature-engineered forecaster with the same model family and tuning budget. Our claim survives only if the latent-state model beats it at sites with unlogged cleanings or long dry spells.
Why not just estimate soiling offline with an existing estimator and feed it in as a feature? Plugging the backward checking estimator's soiling ratio into the forecaster is cheap. Being backward-looking, that estimator needs future clean intervals to calibrate. In real-time use it can supply only the last confirmed value, which is stale exactly after an event. Its rain dependence also breaks down in droughts. We include this two-stage pipeline, restricted to past data only, as a baseline, so the cost of that staleness is measured rather than asserted.
Why not just install cameras, as Ka et al. did? Cameras directly observe soiling and are the most accurate sensor where they exist. They do not scale to millions of rooftops, and the authors themselves flag collection cost. On the CBNU-SolarPower dataset, whose 2,317 images and weather records are the only data here with direct soiling observations, we will ask how much of the camera-driven R² gain a power-only nowcast of the state recovers. The artificial, weekly-placed contaminants have no dynamics to forecast, so this is a nowcasting upper-bound check and not a test of the main claim.
Proposed Experiments
The first experiment validates the state against ground truth. On the US benchmark soiling dataset, whose side-by-side clean and naturally soiled strings in Florida, Oregon, and Colorado give a measured soiling ratio, we compare the online filtered $\tau$ with the measured ratio. Metrics are RMSE and lag in detecting resets, compared against the offline backward checking estimator and Stochastic Rate and Recovery. On the Greece industrial solar park, with seven years of 15-minute SCADA data and 15 logged manual cleanings, we test whether the filter detects cleanings that are absent from its inputs. We withhold the logs and score detection against them. The file reports no numerical RMSE for this dataset, so the comparison here is ours to establish.
The second experiment tests the core claim with day-ahead forecasts on post-event days. On each dataset we select forecast issue times followed by a precipitation, snowfall, or melt event within the horizon. We compare mean bias, RMSE, and pinball loss at the 5%, 50%, and 95% quantiles across the naive, tuned-Kalman, feature-engineered, and full models, plus the weather-independent-hazard ablation. We apply this to Swedish snow seasons using the Checkwatt fleet, where snow is the covering branch and melt the reset. We apply it to the eastern China distributed fleet, where SolarBoost's capacity tracking is the direct incumbent. The decisive pattern is that the full model and the tuned Kalman filter tie on event-free days, while the full model clearly wins on post-event days. That gap should shrink to zero under the ablation.
The third experiment is a smoke stress test. On the NSW wildfire fleet of 160 residential sites, where Ford et al. measured a 13% power reduction per 100 µg/m³ of fine particulates, we test two things. The first is whether particulate-driven deposition in the state explains the part of the loss that persists after air quality recovers. The second is whether fleet pooling sharpens estimates at sites with sparse data. Separating atmospheric dimming from surface deposition is only possible here if residual losses persist after the plume clears. If they do not, we will report that the dataset cannot tell the two apart.
The fourth experiment checks intraday complementarity on the 316-plant US fleet. We measure whether multiplying st-GNN forecasts by our propagated state lowers the 11.01% MAPE at 120 minutes. We expect a small effect at this horizon and report it either way.
Across all experiments, availability of some of these datasets outside their original papers is not established in the file. Where a dataset is unobtainable, we will say so rather than substitute. The proposal fails cleanly if a well-tuned random-walk capacity filter or a feature-engineered tree model matches the full model on post-event days, because that would mean anticipating cleaning adds nothing a simpler model cannot already do.
References
- Solar Power Generation Forecasting via Multimodal Feature Fusion (Student Abstract)
- SolarBoost: Distributed Photovoltaic Power Forecasting Amid Time-varying Grid Capacity
- Data-Driven Soiling Detection in PV Modules
- Snow-Induced PV Loss Modeling Using Production-Data Inferred PV System Models
- Spatiotemporal Graph Neural Network for Performance Prediction of Photovoltaic Power Systems
- Integrating Weather Foundation Model and Satellite to Enable Fine-Grained Solar Irradiance Forecasting
- Prospects of Solar Energy in Rural Areas
- Quantifying the impact of wildfire smoke on solar photovoltaic generation in Australia