Evolutionary AI Model Tames Power Plant Sensor Data With Fewer Errors

A newly published study in Discover Artificial Intelligence proposes a machine learning framework designed specifically for this problem, combining a state space neural operator with genetic algorithm optimization to forecast power plant behavior more accurately, faster and with less memory than established recurrent and attention-based models.

The framework, named GA-SSNOOP for Genetic Algorithm–Optimized State Space Neural Operator, was developed by Jinyong Bai, Jin Bai, Qiang Zhao, Shihan Dang and colleagues. It targets the data middle platforms that modern power plants use to gather, prepare and analyze operational data for decision-making. According to the authors, widely used recurrent and attention-based architectures carry high computational cost, are sensitive to noise and load variations, and in the attention-based case can suffer from catastrophic forgetting. The proposed approach instead represents plant dynamics as continuous-time state evolution, learned through neural operators and tuned by an evolutionary search.

The team evaluated the framework on a Steam Turbine and Boiler Power Plant Dataset containing approximately 50,000 time-series records with 15 to 20 correlated continuous sensor variables, including steam temperature, boiler pressure, turbine inlet conditions, fuel flow rate and efficiency-related measurements. The data were split 70:15:15 into training, validation and testing sets, with missing values handled by interpolation and features with more than ten percent missing entries removed. Outliers were identified using the interquartile range criterion and corrected through interpolation-based replacement.

Architecturally, GA-SSNOOP merges two ideas. A state space module models temporal dynamics with a latent dimension of 64, updating a hidden state through a learned nonlinear transition, with a low-rank constraint on the transition operator to keep the representation compact. A neural operator component, built as a Fourier Neural Operator with four spectral convolution layers of 128 hidden channels and the top 16 Fourier modes retained in each spatial direction, learns mappings between function spaces and remains invariant to the resolution of the underlying data. GELU activations are used in hidden layers, with a linear output layer for continuous regression predictions.

The genetic algorithm layer is what gives the framework its name. Hyperparameters including learning rate, hidden dimension, spectral truncation size, regularization weight and state-space dimension are encoded as normalized chromosome vectors, and an evolutionary process of selection, crossover and mutation searches for configurations that balance prediction accuracy, stability and computational efficiency. The reported optimization used a population of 50, crossover probability of 0.8 and mutation probability of 0.1, terminating after 100 generations or when no improvement occurred for 10 consecutive generations. A multi-objective loss function jointly minimizes prediction error, model complexity, temporal instability and output variability.

The reported results are striking. The model converged within 16 epochs, compared with 22 for the ResDPNN baseline, 25 for NNN-TSFA and 35 for a Transformer model, and achieved a root mean square error of 0.041 and a mean absolute error of 0.032. Across ten independent trials reported as mean plus or minus standard deviation, GA-SSNOOP reached an RMSE of 0.041 plus or minus 0.002, MAE of 0.032 plus or minus 0.001, inference latency of 8.7 plus or minus 0.5 milliseconds, memory usage of 346 plus or minus 12 megabytes, and convergence within 16 plus or minus 1 epochs. A paired t-test with p below 0.05 supported the improvements over baselines.

Efficiency gains were also substantial. The authors report that GA-driven hyperparameter adjustment and the resolution-invariant operator formulation reduce baseline model training time by 34 percent, and that the state-space and neural operator representations cut memory consumption by 29 percent relative to conventional attention-based models. The framework also maintained less than 14 percent performance degradation under sensor noise conditions of 20 to 35 dB signal-to-noise ratio, which the authors contrast with more than 25 percent degradation for adaptive PID and fuzzy synchronization frameworks under equivalent conditions. Settling time under plus or minus 40 percent step load variation decreased by 31 percent, and overshoot fell by 35 percent.

Ablation experiments isolated the contribution of each component. The state space neural operator alone achieved an MAE of 1.68, an RMSE of 2.14 and a forecasting accuracy of 89.72 percent. Adding genetic algorithm optimization reduced MAE to 1.47 and RMSE to 1.88, while the full multi-objective strategy delivered the best values of 1.31 and 1.67. Sensitivity analyses over state dimensions of 32, 64, 128 and 256 identified 64 as the best balance between accuracy and computational cost, and spectral resolutions of 8, 16, 32 and 64 frequency modes showed that 32 offered the best forecasting performance, with no significant gains beyond that point.

Robustness testing extended to realistic industrial disturbances. The team injected synthetic Gaussian noise at levels from 5 to 25 percent of signal variance, simulated sensor degradation at mild, moderate and severe levels through signal drift, measurement bias and random data corruption, and applied controlled load perturbations of plus or minus 20, 40 and 60 percent. The framework maintained stable forecasting accuracy and rapid convergence across these conditions. During inference, peak GPU memory consumption stayed at roughly 1.8 to 2.3 gigabytes and CPU RAM usage at around 3.2 to 4.1 gigabytes, with minimal variation across input sizes. Scaling tests from 10,000 to 100,000 samples raised inference time only from 0.042 to 0.058 seconds per sample.

The authors are candid about limitations. The model assumes a constant temporal sampling rate and may underperform on irregular or missing data. The genetic algorithm requires substantial historical data for initialization, reducing its effectiveness in cold-start or limited-data situations, and evolutionary optimization adds computational overhead when parameter counts are large. The current scheme also emphasizes offline training rather than real-time adaptive learning. Notably, the reported 34 percent convergence-time reduction covers only the training phase after GA-based initialization, with the GA search time reported separately. Future work, the authors write, will pursue Lyapunov-based stability analysis, tighter generalization bounds, federated learning across plants, and cross-domain validation on smart grids, energy storage systems and manufacturing sites.

Subject of Research: A genetic algorithm–optimized state space neural operator framework for forecasting power plant operational time-series data.

Article Title: Genetic algorithm-optimized state space neural operator for power plant data platforms

Article References: Bai, J., Bai, J., Zhao, Q., & Dang, S. (2026). Genetic algorithm-optimized state space neural operator for power plant data platforms. Discover Artificial Intelligence, 6(1), Article 1664. https://doi.org/10.1007/s44163-026-02083-5

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02083-5

Keywords: genetic algorithm, state space neural operator, power plants, time-series forecasting, machine learning, LSTM, Transformer, sensor data, predictive maintenance, industrial analytics, SCADA, deep learning