Sun, Lijia (2018-08). Modeling and Control Techniques in Smart Systems. Doctoral Dissertation. Thesis uri icon

abstract

  • Energy and food crisis are two major problems that our human society has to face in the 21st century. With the world's population reaching 7.62 billion as of May 2018, both electric power and agricultural industries turn to technological innovations for solutions to keep up the increasing demand. In the past and currently, utility companies rely on rule of thumb to estimate power consumption. However, inaccurate predictions often result in over production, and much energy is wasted. On the other hand, traditional periodic and threshold based irrigation practices have also been proven outdated. This problem is further compounded by recent years' frequent droughts across the globe. New technologies are needed to manage irrigations more efficiently. Fortunately, with the unprecedented development of Artificial Intelligence (AI), wireless communication, and ubiquitous computing technologies, high degree of information integration and automation are steadily becoming reality. More smart metering devices are installed today than ever before, enabling fast and massive data collection. Patterns and trends can be more accurately predicted using machine learning techniques. Based on the results, utility companies can schedule production more efficiently, not only enhancing their profitabilities, but also making our world's energy supply more sustainable. In addition, predictions can serve as references to detect anomalous activities like power theft and cyber attacks. On the other hand, with wireless communication, real-time soil moisture sensor readings and weather forecasts can be collected for precision irrigation. Smaller but more powerful controllers provide perfect platforms for complicated control algorithms. We designed and built a fully automated irrigation system at Bushland, Texas. It is designed to operate without any human intervention. Workers can program, move, and monitor multiple irrigation systems remotely. The algorithm that runs on the controls deserves more attention. AI and other state of art controlling techniques are implemented, making it much more powerful than any existing systems. By integrating professional crop yield simulation models like DSSAT, computers can run tens of thousand simulations on all kinds of weather and soil conditions, and more importantly, learn from the experience. In reality, such process would take thousands of years to obtain. Yet, the computers can find an optimum solution in minutes. The experience is then summarized as a policy and stored inside the controller as a lookup table. Furthermore, after each crop season, users can calibrate and update current policy with real harvest data. Crop yield models like DSSAT and AquaCrop play very important roles in agricultural research. They represent our best knowledge in plant biology and can be very accurate when well calibrated. However, the calibration process itself is often time consuming, thus limiting the scale and speed of using these models. We made efforts to combine different models to produce a single accurate prediction using machine learning techniques. The process does not require manual calibration, but only soil, historical weather, and harvest data. 20 models were built, and their results were evaluated and compared. With high accuracy, machine learning techniques have shown a promising direction to best utilize professional models, and demonstrated great potential for use in future agricultural research.
  • Energy and food crisis are two major problems that our human society has to face in the 21st
    century. With the world's population reaching 7.62 billion as of May 2018, both electric power
    and agricultural industries turn to technological innovations for solutions to keep up the increasing
    demand. In the past and currently, utility companies rely on rule of thumb to estimate power
    consumption. However, inaccurate predictions often result in over production, and much energy is
    wasted. On the other hand, traditional periodic and threshold based irrigation practices have also
    been proven outdated. This problem is further compounded by recent years' frequent droughts
    across the globe. New technologies are needed to manage irrigations more efficiently.
    Fortunately, with the unprecedented development of Artificial Intelligence (AI), wireless communication,
    and ubiquitous computing technologies, high degree of information integration and
    automation are steadily becoming reality. More smart metering devices are installed today than
    ever before, enabling fast and massive data collection. Patterns and trends can be more accurately
    predicted using machine learning techniques. Based on the results, utility companies can schedule
    production more efficiently, not only enhancing their profitabilities, but also making our world's
    energy supply more sustainable. In addition, predictions can serve as references to detect anomalous
    activities like power theft and cyber attacks.
    On the other hand, with wireless communication, real-time soil moisture sensor readings and
    weather forecasts can be collected for precision irrigation. Smaller but more powerful controllers
    provide perfect platforms for complicated control algorithms. We designed and built a fully automated
    irrigation system at Bushland, Texas. It is designed to operate without any human intervention.
    Workers can program, move, and monitor multiple irrigation systems remotely. The
    algorithm that runs on the controls deserves more attention. AI and other state of art controlling
    techniques are implemented, making it much more powerful than any existing systems. By integrating
    professional crop yield simulation models like DSSAT, computers can run tens of thousand
    simulations on all kinds of weather and soil conditions, and more importantly, learn from the experience. In reality, such process would take thousands of years to obtain. Yet, the computers can
    find an optimum solution in minutes. The experience is then summarized as a policy and stored
    inside the controller as a lookup table. Furthermore, after each crop season, users can calibrate and
    update current policy with real harvest data.
    Crop yield models like DSSAT and AquaCrop play very important roles in agricultural research.
    They represent our best knowledge in plant biology and can be very accurate when well
    calibrated. However, the calibration process itself is often time consuming, thus limiting the scale
    and speed of using these models. We made efforts to combine different models to produce a single
    accurate prediction using machine learning techniques. The process does not require manual calibration,
    but only soil, historical weather, and harvest data. 20 models were built, and their results
    were evaluated and compared. With high accuracy, machine learning techniques have shown a
    promising direction to best utilize professional models, and demonstrated great potential for use in
    future agricultural research.

publication date

  • August 2018
  • August 2018