Solar farms forecast better when they can see their neighbors. A cloud bank that dims an upwind plant will reach a downwind plant minutes later. A graph neural network, which treats each plant as a node and lets nearby plants share information, turns that warning into a much more accurate forecast. The price is that competing producers must share live generation data they treat as commercially sensitive. Homomorphic encryption lets a server compute on encrypted numbers without seeing them. It removes that price, but only for models built entirely from additions and multiplications, so every nonlinear activation such as ReLU must be replaced by a polynomial. Polynomials are well behaved only on a bounded input range. Outside that range they blow up.
This note argues that solar physics supplies that bounded range for free. Each plant can divide out its own clear-sky curve locally, in plaintext, before encrypting anything. That turns raw power into a dimensionless "clear-sky index" confined to a narrow interval, and all the trigonometry of sun position stays off the encrypted path. A shallow network with quadratic activations, trained from scratch on these bounded residuals, can then run non-interactively on a server within a 15-minute dispatch window. The concrete bet is that such a network recovers most of the known cross-site accuracy gain. The central experimental question is whether its quadratic terms actually beat an encrypted linear cross-site model, which needs no activation approximation at all.
Core Claim
Two direction-specific facts fit together. Work on fully polynomial networks locates the main failure of polynomial activations in inputs drifting out of range: variance grows exponentially with stacked polynomial layers. ULD-Net counters this with per-sample polynomial normalization and variance penalties, but those layers cost encrypted multiplications. Solar forecasting work, meanwhile, already removes the deterministic diurnal arc to leave a stationary residual. The Modified Auto Regressive model does this through "ensemble deduction" (subtracting the historical mean for each time of day). The spatiotemporal GNN of Karimi et al. does it by aligning sites to local solar noon.
Our claim is that doing this normalization per site in plaintext makes the normalization layers inside the encrypted network largely unnecessary, which saves multiplicative depth. It also resolves a tension flagged in the file: sun-geometry features such as the cosine of the zenith angle clash with addition-and-multiplication-only arithmetic. Here they never enter the encrypted computation.
There is also a physical reason to expect degree two to be the right amount of nonlinearity. A useful advection signal is roughly a product: an upwind drop in clear-sky index matters in proportion to how clear the downwind site currently is and how well the pair aligns with the cloud motion. A quadratic activation represents such products directly. Whether that intuition holds is exactly what the linear baseline tests.
Research Program
The deployment setting mirrors the vertical collaboration of pwXGBoost. In that SMPC system, a target wind farm keeps its labels while neighbors contribute features through interactive secret sharing. In our version, the target site publishes a CKKS public key (CKKS is the homomorphic scheme for approximate real-number arithmetic). Each neighbor computes its clear-sky-index history locally and encrypts it under that key. A forecasting server holds plaintext model weights and evaluates the network on the ciphertexts in one non-interactive round. Only the target can decrypt its forecast.
The server learns nothing about the inputs. Neighbors' data reaches the target only through the forecast itself, and we will state that leakage explicitly rather than claim it away. Training is not the focus. We assume the model is fit once on a historical window pooled under contract or trained offline with secure computation, because training is not latency-critical. Encrypted inference is the recurring cost and the contribution.
The network follows Karimi et al.: temporal convolutions interleaved with graph convolutions whose edges are Gaussian kernels of the distance between sites, with small weights zeroed out. Every activation is replaced by a trainable quadratic, as in ULD-Net, and the model is trained from scratch rather than approximated after the fact. Following CryptoGCN, we prune activation layers that contribute little and pack node features along the sparse graph structure to avoid ciphertext rotations. CryptoGCN assumes a public adjacency matrix. That is acceptable here if site locations are public plaintext known to the server; if they are not, the approach does not carry over unchanged.
Horizons from 15 to 120 minutes are produced by separate direct output heads rather than by feeding predictions back in. This choice keeps the circuit depth fixed and avoids the untested variance feedback that recursive rollouts introduce. We will separately measure what happens under recursive rollout to probe that feedback directly.
We analyze three things. The first is the multiplicative-depth budget: every plaintext-weight multiply and every square consumes a rescale level, and the target is a circuit that needs no bootstrapping. The second is the empirical distribution of pre-activation values across seasons and storm days, measured against the interval on which the quadratics were trained. The third is how often anomalous days, such as the "overnight shock" failures reported for differential polynomial networks on PV, push activations out of range. We also give the method a second, wind-oriented test. Wind power follows an approximately cubic power curve, so a degree-3 activation is the physically natural choice there. This test runs wherever wind telemetry with neighbors is available.
Baselines
The do-nothing reference is clear-sky persistence at each site, which assumes the current clear-sky index holds for the whole horizon. It costs nothing and needs no sharing. No reported result was found on Lacuna for clear-sky persistence on the 316-PV dataset, so we will compute it ourselves.
The local-only reference is a temporal model with no neighbor data. On the 316-PV dataset, Karimi et al. report 18.98% MAPE at 120 minutes for their temporal-only 1-D convolution. We also rerun the MAR ensemble-deduction model, which was competitive with CNN and LSTM models at short horizons on tropical data. Together these define the gain that privacy-preserving collaboration has to buy.
The strongest existing method is the plaintext spatiotemporal GNN, at 11.01% MAPE at 120 minutes on the same dataset (Karimi et al.). It requires full data sharing, so it serves as the accuracy ceiling rather than a competitor that respects the privacy constraint. We retrain it with our clear-sky preprocessing and our splits so that all numbers are comparable.
The strongest privacy-preserving competitor is interactive SMPC boosting, pwXGBoost. On Inner Mongolia wind telemetry at a 4-hour horizon, it reports 3.781% RMSE against 4.033% for local XGBoost, with inference under 400 seconds. We reimplement it on the PV data with the same neighbor sets and the same tuning budget.
The critical ablation is an encrypted linear cross-site model: the same graph and temporal structure with degree-1 activations, which is equivalent to a vector autoregression. It needs no polynomial approximation at all. A second ablation is the post-hoc route: train a ReLU GNN, then substitute Chebyshev fits of increasing degree. A third feeds raw rather than clear-sky-normalized inputs, with and without per-sample normalization layers and ULD-Net variance penalties. This third ablation isolates whether physics-based bounding is what makes the low-degree network stable.
All models get identical data, identical hyperparameter search budgets, and the same hardware for encrypted timing. Beating the field means three things at once. The encrypted quadratic model must close at least most of the 18.98-to-11.01 gap. It must beat the encrypted linear model by a margin larger than seed-to-seed variation at 60 to 120 minutes. And it must match or beat pwXGBoost accuracy with one communication round.
Why not just…?
Why not just use an encrypted linear model and skip activation approximation entirely? This is the most serious threat. A linear model runs cheaply under homomorphic encryption, and linear autoregression on deseasonalized residuals is known to be competitive at single sites (MAR). Karimi et al.'s gain came from adding space, not from proving nonlinearity was necessary. If the degree-1 ablation lands within noise of the quadratic model, the correct conclusion is "deploy encrypted linear cross-site regression." We would report that result rather than dress up the quadratic. We expect the gap to appear mainly on ramp events, where the gating product described above matters, so we evaluate ramp-conditioned error separately.
Why not just use SMPC as pwXGBoost does? It already works and handles comparisons natively. However, it is interactive: every nonlinear step costs network round trips among competing parties. MPCFormer shows that replacing such nonlinearities with quadratics gave 2.2× to 5.9× speedups under MPC, which suggests the round-trip cost dominates. Our comparison measures accuracy, wall-clock time, and bytes exchanged against pwXGBoost on identical neighbor sets.
Why not just train on differentially private or federated updates? Noise calibrated to hide individual generation curves fights directly against short-horizon ramp detection, the one signal neighbors are valuable for. Federated training also does not protect live inputs at inference time, which is when the commercially sensitive real-time data flows. We do not run a differential-privacy baseline. That is a limitation.
Why not just fit high-degree polynomials to a trained ReLU network? PUMA shows that piecewise approximations can preserve accuracy without retraining in MPC. Under homomorphic encryption, though, ULD-Net reports that training quadratics from scratch beat a degree-81 prior on ImageNet at 69.79% top-1 with a 2.76× speedup. Our post-hoc ablation checks whether that ordering holds on bounded, forecasting-type inputs, where post-hoc fitting might fare better than on images.
Proposed Experiments
The primary benchmark is the 316-PV-system dataset: 15-minute resolution, two years, solar-noon aligned. We use chronological splits and report MAPE and RMSE at 15, 30, 60, and 120 minutes, plus error conditioned on ramp events. Encrypted evaluation reports multiplicative depth, whether bootstrapping was needed, latency per forecast round, and ciphertext bytes per site. We also report the maximum deviation between encrypted and plaintext outputs, which should sit at the level of CKKS numerical precision.
We rerun the scaling study from 77 to 316 sites, because encrypted cost grows with neighbor count while accuracy gains may saturate. A wind replication with degree-3 activations follows if the Inner Mongolia telemetry or a comparable multi-farm set can be obtained. Otherwise, the wind claim remains untested.
Three checks decide the outcome. First, the fraction of the local-to-plaintext-GNN gap recovered by the encrypted quadratic model. Second, the margin over the encrypted linear model on ramp events. Third, the fraction of stormy or anomalous test days on which pre-activations leave the trained interval, with and without clear-sky normalization; the physics-bounding claim fails if normalization does not sharply reduce that fraction. Finally, a recursive-rollout variant measures how activation variance grows with rollout steps, giving the first direct evidence on whether direct multi-horizon heads are necessary.
References
- ULD-Net: Enabling Ultra-Low-Degree Fully Polynomial Networks for Homomorphically Encrypted Inference
- Modified Auto Regressive Technique for Univariate Time Series Prediction of Solar Irradiance
- Spatiotemporal Graph Neural Network for Performance Prediction of Photovoltaic Power Systems
- Privacy Preserving Ultra-Short-term Wind Power Prediction Based on Secure Multi Party Computation
- CryptoGCN: Fast and Scalable Homomorphically Encrypted Graph Convolutional Network Inference
- Modelling the Photovoltaic output Power using the differential Polynomial Network and Evolutionary fuzzy Rules
- MPCFORMER: FAST, PERFORMANT AND PRIVATE TRANSFORMER INFERENCE WITH MPC
- PUMA: Secure Inference of LLaMA-7B in Five Minutes