Buildings are caught in a persistent tug-of-war between light and heat. The larger the windows, the more daylight floods a room and the less electricity is spent on artificial lighting, but the same glass that welcomes the sun also admits solar heat, driving up summer cooling loads and winter heat losses. A new study published in Discover Artificial Intelligence by Yi Mei of Chongqing Metropolitan College of Science and Technology tackles this dilemma head-on with a computational framework that optimizes a building’s thermal and light environment simultaneously, then keeps it under intelligent real-time control. The results are striking: an energy saving rate of 17.4 percent, an 8.2 percent drop in thermal dissatisfaction, a 21.1 percent increase in illumination compliance, and an overall jump in user satisfaction from 68.2 percent to 86.3 percent.
The core of the research is a multi-objective optimization model that weighs four competing goals at once: energy consumption, thermal comfort, visual comfort, and full life cycle cost. Thermal comfort is quantified through the Predicted Percentage of Dissatisfied occupants, or PPD, which is derived from the Fanger PMV index, a heat-balance equation that accounts for metabolic rate, clothing insulation, air temperature, radiant temperature, humidity, and air velocity. Visual comfort is measured by the Daylight Factor, the ratio of indoor to outdoor natural illumination under an overcast sky, while glare is policed by the Unified Glare Rating, which the study treats as a hard constraint capped at 16 rather than a soft objective. Crucially, the model recognizes that these objectives are physically coupled: the window-to-wall ratio and the solar heat gain coefficient jointly determine both how much sunlight enters a room and how much heat rides in with it, creating a nonlinear trade-off that no single-parameter design can resolve.
To capture that coupling, the study builds a predictive engine based on an improved Genetic Algorithm-Back Propagation neural network, or GA-BP. A genetic algorithm scours the space of initial network weights and thresholds to find a favorable starting point, reducing the risk of getting trapped in local minima, and then backpropagation fine-tunes those parameters through error-driven learning. Two refinements set this version apart from the standard recipe. First, the random seed of the Mersenne Twister generator is fixed at 1, so every training run begins from identical initial conditions, making results reproducible and eliminating random fluctuations that would otherwise muddy the analysis of which design variables matter most. Second, instead of a fixed training-test split, the researchers employ a metabolic dataset partitioning method: samples are ordered chronologically, and at each prediction step only the preceding observations train the model while the next observation is reserved exclusively for testing. The window then slides forward one step at a time, so the network is always learning from the freshest data, which improves its adaptability to short-term trends and sudden changes such as heating load swings.
The payoff in prediction accuracy is substantial. The improved GA-BP achieved a coefficient of determination of 0.961, well above the 0.895 of the standard GA-BP, and its root mean square error of 0.174 undercut ARIMA, support vector regression, and even long short-term memory networks. In multi-step forecasting, the mean absolute error rose only from 0.103 at one hour to 0.158 at 24 hours, indicating that the model remains stable over a full-day prediction horizon, exactly what a predictive controller needs to plan ahead.
On top of this predictor sits the neural network model predictive control framework, or NNMPC. Rather than reacting to threshold violations like a conventional thermostat, NNMPC looks into the future: at each sampling instant, the GA-BP model forecasts indoor temperature and illuminance over a prediction horizon, and the controller solves a constrained optimization problem to find the best sequence of control actions, which include air-conditioning supply temperature and lighting power. Only the first action in the sequence is executed; then new sensor measurements arrive, the horizon rolls forward, and the optimization repeats. This receding-horizon mechanism, weighted across energy consumption, thermal deviation, and illuminance deviation, forms a closed loop of state sensing, prediction and optimization, control execution, and feedback correction.
Finding the optimal static design configuration is the job of a parallel hybrid optimizer that fuses NSGA-II, a genetic algorithm strong at global exploration but slow to converge, with MOPSO, a particle swarm method that converges quickly but can stagnate in local optima. In each iteration, both algorithms generate 50 offspring from a population of 100, and fast non-dominated sorting selects the next generation while duplicate Pareto solutions are pruned to preserve diversity. The hybrid’s Inverted Generational Distance, a measure of how close and evenly spread the solutions are relative to the true Pareto front, reached 0.039, more than 40 percent lower than either single algorithm, and it converged after just 118 iterations compared with 156 for NSGA-II and 143 for MOPSO. The computational cost of 263.5 seconds is acceptable because this optimization runs offline; real-time decisions are left to NNMPC.
The case study applied the framework to a five-story middle school teaching building in Tianjin, China, with a gross floor area of 4,532.41 square meters, using typical meteorological year data from the Beijing region of the China Standard Weather Database. Eight design variables were optimized, including exterior wall and roof insulation thicknesses and window parameters. Among the Pareto solutions, the best energy design achieved 60.53 kilowatt-hours per square meter per year, the best thermal comfort solution cut dissatisfaction to 22.85 percent, and the best daylighting solution reached a daylight factor of 3.42 percent, though with glare near its limit. A cost-priority design delivered the largest life-cycle savings of 479,138 yuan but sacrificed daylighting, dropping the daylight factor from 1.85 percent to 1.52 percent, a clear illustration of the trade-off between economics and visual comfort. The comprehensive balanced solution, which performed at a medium-to-optimal level across all four objectives, emerged as the most practical choice for conventional buildings.
The dynamic control layer then proved its worth in simulation and in the field. Under NNMPC, indoor temperature held near 18 degrees Celsius and 300 lux during working hours on a typical day, and on a hot summer day the system kept interiors at 26 degrees even when outdoor temperatures exceeded 30. Compared with traditional threshold-based PI control, the sensor-linked lighting and air-conditioning strategy cut total annual energy consumption from 73.11 to 60.37 kilowatt-hours per square meter per year, saved 52,400 yuan in life-cycle cost, and lifted user satisfaction from 68.2 to 86.3 percent. Field verification confirmed the simulation: indoor temperature deviation shrank from plus or minus 1.3 degrees to plus or minus 0.6 degrees, illumination deviation fell by 55.3 percent, and HVAC and lighting energy use dropped by 14.7 percent and 19.9 percent respectively.
Robustness testing across climates revealed both strengths and limits. In Tianjin, whose climate resembles Beijing’s, energy, comfort, and cost indicators shifted by at most about 1.7 percent. But in Shanghai’s hot, humid summers the energy loss rate reached 8.1 percent, and in Guangzhou, with longer and more intense cooling seasons, it climbed to 14.9 percent, showing that the Beijing-trained model generalizes well to adjacent climates but struggles where thermal and solar conditions diverge sharply. The author suggests that climate-specific transfer learning or online model updating could close that gap, along with small-sample learning techniques for new buildings that lack the historical data on which prediction accuracy depends.
The broader significance is architectural as much as computational. By embedding a fast, reproducible neural predictor inside a rolling-horizon controller, the study demonstrates that the century-old conflict between daylight and heat need not be settled once at the drawing board and then endured forever. Instead, the envelope can be optimized offline for lifetime economics and comfort, while sensors and a predictive brain continuously renegotiate the balance between sunlight, lamps, and air conditioning as weather and occupancy shift hour by hour. For school buildings, where poor lighting has been linked to rising myopia rates and poor thermal conditions to reduced learning efficiency, the 18-point gain in user satisfaction may prove the most consequential number of all.
Subject of Research: Multi-objective optimization and neural network model predictive control of indoor thermal and lighting environments in buildings
Article Title: Multi-objective intelligent optimization and control of indoor thermal and light environment based on NNMPC
Article References: Mei, Y. (2026). Multi-objective intelligent optimization and control of indoor thermal and light environment based on NNMPC. Discover Artificial Intelligence, 6(1), Article 1216. https://doi.org/10.1007/s44163-026-02299-5
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02299-5
Keywords: NNMPC, GA-BP neural network, multi-objective optimization, thermal comfort, daylighting, building energy efficiency, NSGA-II, MOPSO, model predictive control, smart buildings, life-cycle cost, sensor-based control
News Source: Cassandra Pierce. (October 19, 2026). Neural Network Predictive Control Cuts Building Energy Use While Boosting Comfort. Scienmag.
