Hybrid Cross-Attention Revision for Ultra-Short-Term Photovoltaic Forecasting

Photovoltaic output departs from the weather-model expectation when sub-grid atmospheric processes and initial-condition errors distort the local irradiance history that actually governs panel response over the next few hours.   Grid operation needs forecasts that are short enough to reflect rapid meteorological change and stable enough to support dispatch, yet the data streams used for prediction do not carry the same kind of information. Numerical weather prediction supplies forward-looking meteorological sequences. Station instruments, on the other hand, record the local radiative and meteorological state directly, along with the plant’s power output, but they do not provide the future observations needed for actual forecasting. The central issue is not data quantity. It is that future model fields and past station measurements describe the same plant state on different time bases and from different observational standpoints. In a recent research paper published in International Journal of Electrical Power & Energy Systems, Professor Yuanbing Wang’s team from the Nanjing University of Information Science and Technology, developed a two-stage ultra-short-term photovoltaic forecasting framework that couples a CNN-BiLSTM short-term predictor with a CNN-Transformer revision model using cross-attention. The paper’s first author is student Yuchen Dai, with Professor Wang as the corresponding author. The research team also includes Professor Yaodeng Chen and students Junwen Wu and Jie Chao. Many plants already collect these station observations routinely, so the forecasting task can be recast around using them to revise a near-future sequence generated from weather-model input and in that setup, historical local measurements do not remain passive training data; they directly participate in correcting the short-term forecast.

They first used future meteorological model data in 24-hour sliding windows of 96 points to train a CNN-BiLSTM single-step predictor, producing a short-term output sequence for the coming day. They then extracted the first 4 hours of that forecast and aligned it with two station-side histories from the preceding 4 hours: measured meteorological observations and measured photovoltaic output.  A CNN-BiLSTM pathway suits the first stage because the convolutional layers capture local sequence structure and the bidirectional recurrent layer preserves dependence across the 24-hour input. The Transformer serves a different role in the second stage, where cross-attention lets the model treat the short-term forecast as a sequence to be revised using recent station history as the conditioning context. In the cross-attention block, the short-term prediction supplies the value vector, and the station-side histories supply query and key, so revision is driven by how the local past reweights the model-based future.

The source reports actual 15-minute data from four Jiangsu plants spanning April to December 2023, paired with ECMWF forecast data on the same interval. The team cleaned plant output with support vector regression using surface solar radiation as a reference variable and removed points above an empirically chosen error threshold of 80 W/(m²·sr). They also introduced a time-ratio feature marking each observation’s position within the daily cycle. That step matters because photovoltaic production follows not only weather variation but also the daily light cycle. Encoding time explicitly prevented anomalous nighttime generation in prediction and sharpened the periodic form of the output sequence. The sliding-window construction then preserved local temporal continuity for both the 24-hour and 4-hour problems, letting the two models operate on sequences built from the plant’s actual operating rhythm rather than on isolated samples.

The study contrasts five settings: a station-only 24-hour CNN-BiLSTM, a model-only 24-hour CNN-BiLSTM, a weighted fusion of those two outputs, a 4-hour CNN-BiLSTM built from matched station-side and model-side data, and the full hybrid 4-hour method. Averaged across the four stations, the proposed model reached an MAE of 3.41 and an RMSE of 6.21, compared with 4.09 and 8.57 for the model-only 24-hour baseline. That corresponds to reductions of about 16.7% in MAE and 27.4% in RMSE. Relative to output weighted fusion, the reductions were about 16.8% and 27.9%; relative to the 4-hour CNN-BiLSTM, they were about 7.7% and 20.1%. Station-level results followed the same pattern: MAE and RMSE fell by about 16.22% and 27.42% at Station A, 23.29% and 33.20% at Station B, 20.38% and 29.34% at Station C, and 6.83% and 14.59% at Station D. The paper further notes that the hybrid forecasts tracked peaks more closely under stable weather, followed hourly variation with greater detail, and improved the timing of sunrise and sunset transitions. Peak shape and day-edge timing directly affect any calculation that integrates power over time, so better local alignment changes the practical content of the forecast itself. After the researchers computed Pearson correlations among meteorological variables and power output, they built four groups, starting with the two most correlated radiation variables and then expanding with clouds, temperature, visibility, wind, and additional radiation terms. Group 3, which combined ssrd, ssr, tsr, tsrc, ssrc, fdir, cdir, mcc, tcc, t2m, vis, ws, and wd, produced the best ultra-short-term metrics, with MAE 1.3492 and RMSE 3.0049. The new work interprets this result by showing that strongly positive radiation variables contributed most effectively during ultra-short-term revision, where they shaped the modal changes of the forecast sequence. Fine-grained radiative detail becomes most useful after the forecast has already been formed, when the model needs to adjust short-horizon dynamics rather than build the entire day from scratch.

The new research paper changes the forecasting logic from direct prediction to forecast revision driven by recent plant-specific evidence. In many photovoltaic workflows, meteorological forecast data serve as the main predictive substrate, while station observations remain historical inputs. Here, the short-term forecast becomes an intermediate physical statement rather than a terminal answer. Ultra-short-term prediction then becomes a correction problem in which recent plant behavior reshapes the near-future sequence generated from weather-model input. The CNN-BiLSTM stage compresses and propagates temporal structure from future meteorological model input across a 24-hour horizon. Its cross-attention mechanism lets local station histories interrogate the already-formed forecast. It expresses two distinct sources of information in photovoltaic prediction: the broad meteorological trajectory supplied by the forecast model and the near-field operating state recorded at the plant. Keeping those roles distinct makes the revision interpretable in functional terms. The plant-side sequence is not asked to predict the future by itself, which would conflict with the absence of future observations. Instead, it modifies a forecast that already exists, and that is precisely the regime where local measurements carry the most value.

Radiation variables with strong positive correlation to power output, joined with cloud, temperature, visibility, and wind information, proved especially effective in the revision stage. The paper’s own interpretation is that these variables enrich the dynamic response of the ultra-short-term forecast. That matters because photovoltaic output is shaped not only by bulk irradiance magnitude but also by how radiative and meteorological factors vary over short intervals. This framing assigns short-scale radiative variation a distinct role during forecast revision rather than during initial sequence construction. The study, then, contributes a way of thinking about feature utility: some variables may be most informative not at the stage of initial day-ahead sequence construction, but at the stage where short-horizon corrections are made to evolving local behavior. The study also frames the method as portable. It states that deployment in a new region requires three inputs: historical weather observations, historical photovoltaic generation data, and numerical weather prediction fields obtainable from geographic coordinates, with adaptation handled through parameter fine-tuning rather than architectural redesign.   A method that depends on widely available plant records and standard forecast products fits naturally into operating environments where bespoke instrumentation is not the main route to better forecasting. What the study contributes, then, is a hybrid revision framework for ultra-short-term photovoltaic prediction in which recent station evidence and forecast-model sequences are not competing alternatives. They become coordinated layers of information, each used at the stage where it carries the greatest explanatory force for the next few hours of plant output.